Daily Papers

Large language models (LLMs) have played a fundamental role in various natural language processing tasks with powerful prompt techniques. However, in real-world applications, there are often similar prompt components for repeated queries, which causes significant computational burdens during inference. Existing prompt compression and direct fine-tuning methods aim to tackle these challenges, yet they frequently struggle to strike an optimal balance between cost-efficiency and performance effectiveness, especially in complex tasks such as NL2Code. In this paper, we propose a novel method namely PromptIntern to internalize the prompt knowledge into model parameters via progressive fine-tuning. Our method enables LLMs to emulate the human learning process for a new task, where detailed templates and examples in a prompt are gradually internalized and phased out progressively as the model grows accustomed to the task. Extensive experiments demonstrate that our method reduces inference tokens over 90%, speedups inference by 4.2 times, and saves 88.3% monetary cost.

Post-training is essential for adapting Large Language Models (LLMs) to real-world applications. Deploying post-trained models faces significant challenges due to substantial memory overhead and noticeable inference latency. Existing work has identified significant redundancies in LLMs and proposed efficient architectures, namely intra-layer KV sharing and cross-layer KV sharing. However, intra-layer KV sharing still results in high inference costs, while cross-layer KV sharing leads to significant performance degradation. As a result, both methods remain suboptimal for post-training pre-trained LLMs. In this paper, we identify that the Softmax operation is a primary bottleneck for LLM inference and discover that it is actually highly redundant during post-training. We propose Softmax Unification in Attention (UniAttn), a novel post-training method that unifies Softmax activations across transformer blocks to reduce LLM inference costs. Additionally, UniAttn adopts a linear projection to compensate for the errors induced by Softmax unification. Experiments show that UniAttn matches the performance of standard post-training while significantly reducing inference costs, outperforming existing efficient architectures during post-training. Our code will be available at https://github.com/Bostoncake/UniAttn.

With the fast growth of parameter size, it becomes increasingly challenging to deploy large generative models as they typically require large GPU memory consumption and massive computation. Unstructured model pruning has been a common approach to reduce both GPU memory footprint and the overall computation while retaining good model accuracy. However, the existing solutions do not provide a highly-efficient support for handling unstructured sparsity on modern GPUs, especially on the highly-structured Tensor Core hardware. Therefore, we propose Flash-LLM for enabling low-cost and highly-efficient large generative model inference with the sophisticated support of unstructured sparsity on high-performance but highly restrictive Tensor Cores. Based on our key observation that the main bottleneck of generative model inference is the several skinny matrix multiplications for which Tensor Cores would be significantly under-utilized due to low computational intensity, we propose a general Load-as-Sparse and Compute-as-Dense methodology for unstructured sparse matrix multiplication. The basic insight is to address the significant memory bandwidth bottleneck while tolerating redundant computations that are not critical for end-to-end performance on Tensor Cores. Based on this, we design an effective software framework for Tensor Core based unstructured SpMM, leveraging on-chip resources for efficient sparse data extraction and computation/memory-access overlapping. At SpMM kernel level, Flash-LLM significantly outperforms the state-of-the-art library, i.e., Sputnik and SparTA by an average of 2.9x and 1.5x, respectively. At end-to-end framework level on OPT-30B/66B/175B models, for tokens per GPU-second, Flash-LLM achieves up to 3.8x and 3.6x improvement over DeepSpeed and FasterTransformer, respectively, with significantly lower inference cost.

Since ChatGPT released its API for public use, the number of applications built on top of commercial large language models (LLMs) increase exponentially. One popular usage of such models is leveraging its in-context learning ability and generating responses given user queries leveraging knowledge obtained by retrieval augmentation. One problem of deploying commercial retrieval-augmented LLMs is the cost due to the additionally retrieved context that largely increases the input token size of the LLMs. To mitigate this, we propose a token compression scheme that includes two methods: summarization compression and semantic compression. The first method applies a T5-based model that is fine-tuned by datasets generated using self-instruct containing samples with varying lengths and reduce token size by doing summarization. The second method further compresses the token size by removing words with lower impact on the semantic. In order to adequately evaluate the effectiveness of the proposed methods, we propose and utilize a dataset called Food-Recommendation DB (FRDB) focusing on food recommendation for women around pregnancy period or infants. Our summarization compression can reduce 65% of the retrieval token size with further 0.3% improvement on the accuracy; semantic compression provides a more flexible way to trade-off the token size with performance, for which we can reduce the token size by 20% with only 1.6% of accuracy drop.

Vision Language Models (VLMs) have demonstrated strong capabilities across various visual understanding and reasoning tasks. However, their real-world deployment is often constrained by high latency during inference due to substantial compute required to process the large number of input tokens (predominantly from the image) by the LLM. To reduce inference costs, one can either downsize the LLM or reduce the number of input image-tokens, the latter of which has been the focus of many recent works around token compression. However, it is unclear what the optimal trade-off is, as both the factors directly affect the VLM performance. We first characterize this optimal trade-off between the number of visual tokens and LLM parameters by establishing scaling laws that capture variations in performance with these two factors. Our results reveal a surprising trend: for visual reasoning tasks, the inference-optimal behavior in VLMs, i.e., minimum downstream error at any given fixed inference compute, is achieved when using the largest LLM that fits within the inference budget while minimizing visual token count - often to a single token. While the token reduction literature has mainly focused on maintaining base model performance by modestly reducing the token count (e.g., 5-10times), our results indicate that the compute-optimal inference regime requires operating under even higher token compression ratios. Based on these insights, we take some initial steps towards building approaches tailored for high token compression settings. Code is available at https://github.com/locuslab/llava-token-compression.

We investigate an emerging setup in which a small, on-device language model (LM) with access to local data communicates with a frontier, cloud-hosted LM to solve real-world tasks involving financial, medical, and scientific reasoning over long documents. Can a local-remote collaboration reduce cloud inference costs while preserving quality? First, we consider a naive collaboration protocol where the local and remote models simply chat back and forth. Because only the local model reads the full context, this protocol achieves a 30.4x reduction in remote costs, but recovers only 87% of the performance of the frontier model. We identify two key limitations of this protocol: the local model struggles to (1) follow the remote model's multi-step instructions and (2) reason over long contexts. Motivated by these observations, we study an extension of this protocol, coined MinionS, in which the remote model decomposes the task into easier subtasks over shorter chunks of the document, that are executed locally in parallel. MinionS reduces costs by 5.7x on average while recovering 97.9% of the performance of the remote model alone. Our analysis reveals several key design choices that influence the trade-off between cost and performance in local-remote systems.

Large language models have emerged as a versatile tool but are challenging to apply to tasks lacking large inference budgets and large in-domain training sets. This work formalizes these constraints and distinguishes four important variables: the pretraining budget (for training before the target domain is known), the specialization budget (for training after the target domain is known), the inference budget, and the in-domain training set size. Across these settings, we compare different approaches from the machine learning literature. Limited by inference cost, we find better alternatives to the standard practice of training very large vanilla transformer models. In particular, we show that hyper-networks and mixture of experts have better perplexity for large pretraining budgets, while small models trained on importance sampled datasets are attractive for large specialization budgets.

Transformer-based large Language Models (LLMs) become increasingly important in various domains. However, the quadratic time complexity of attention operation poses a significant challenge for scaling to longer contexts due to the extremely high inference latency and GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to accelerate attention computation. To leverage the dynamic sparse property of attention, RetrievalAttention builds approximate nearest neighbor search (ANNS) indexes upon KV vectors in CPU memory and retrieves the most relevant ones via vector search during generation. Due to the out-of-distribution (OOD) between query vectors and key vectors, off-the-shelf ANNS indexes still need to scan O(N) (usually 30% of all keys) data for accurate retrieval, which fails to exploit the high sparsity. RetrievalAttention first identifies the OOD challenge of ANNS-based attention, and addresses it via an attention-aware vector search algorithm that can adapt to queries and only access 1--3% of data, thus achieving a sub-linear time complexity. RetrievalAttention greatly reduces the inference cost of long-context LLM with much lower GPU memory requirements while maintaining the model accuracy. Especially, RetrievalAttention only needs 16GB GPU memory for serving 128K tokens in LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds on a single NVIDIA RTX4090 (24GB).

The widespread adoption of AI systems in the economy hinges on their ability to generate economic value that outweighs their inference costs. Evaluating this tradeoff requires metrics that account for both performance and costs. We propose a framework grounded in production theory for evaluating language models by combining accuracy and inference cost. We introduce "cost-of-pass", the expected monetary cost of generating a correct solution. We then define the "frontier cost-of-pass" as the minimum cost-of-pass achievable across available models or the "human-expert, using the approximate cost of hiring an expert. Our analysis reveals distinct economic insights. First, lightweight models are most cost-effective for basic quantitative tasks, large models for knowledge-intensive ones, and reasoning models for complex quantitative problems, despite higher per-token costs. Second, tracking this frontier cost-of-pass over the past year reveals significant progress, particularly for complex quantitative tasks where the cost has roughly halved every few months. Third, to trace key innovations driving this progress, we examine counterfactual frontiers: estimates of cost-efficiency without specific model classes. We find that innovations in lightweight, large, and reasoning models have been essential for pushing the frontier in basic quantitative, knowledge-intensive, and complex quantitative tasks, respectively. Finally, we assess the cost-reductions afforded by common inference-time techniques like majority voting and self-refinement, finding that their marginal accuracy gains rarely justify their costs. Our findings underscore that complementary model-level innovations are the primary drivers of cost-efficiency, and our economic framework provides a principled tool for measuring this progress and guiding deployment.

Large Language Models (LLMs) have a natural role in answering complex queries about data streams, but the high computational cost of LLM inference makes them infeasible in many such tasks. We propose online cascade learning, the first approach to address this challenge. The objective here is to learn a "cascade" of models, starting with lower-capacity models (such as logistic regression) and ending with a powerful LLM, along with a deferral policy that determines the model to be used on a given input. We formulate the task of learning cascades online as an imitation-learning problem, where smaller models are updated over time imitating the collected LLM demonstrations, and give a no-regret algorithm for the problem. Experimental results across four benchmarks show that our method parallels LLMs in accuracy while cutting down inference costs by as much as 90% with strong robustness against input distribution shifts, underscoring its efficacy and adaptability in stream processing.

Transformer models, despite their impressive performance, often face practical limitations due to their high computational requirements. At the same time, previous studies have revealed significant activation sparsity in these models, indicating the presence of redundant computations. In this paper, we propose Dynamic Sparsified Transformer Inference (DSTI), a method that radically reduces the inference cost of Transformer models by enforcing activation sparsity and subsequently transforming a dense model into its sparse Mixture of Experts (MoE) version. We demonstrate that it is possible to train small gating networks that successfully predict the relative contribution of each expert during inference. Furthermore, we introduce a mechanism that dynamically determines the number of executed experts individually for each token. DSTI can be applied to any Transformer-based architecture and has negligible impact on the accuracy. For the BERT-base classification model, we reduce inference cost by almost 60%.

Large language models (LLMs) are central to modern natural language processing, delivering exceptional performance in various tasks. However, their intensive computational and memory requirements present challenges, especially for devices with limited DRAM capacity. This paper tackles the challenge of efficiently running LLMs that exceed the available DRAM capacity by storing the model parameters on flash memory but bringing them on demand to DRAM. Our method involves constructing an inference cost model that harmonizes with the flash memory behavior, guiding us to optimize in two critical areas: reducing the volume of data transferred from flash and reading data in larger, more contiguous chunks. Within this flash memory-informed framework, we introduce two principal techniques. First, "windowing'" strategically reduces data transfer by reusing previously activated neurons, and second, "row-column bundling", tailored to the sequential data access strengths of flash memory, increases the size of data chunks read from flash memory. These methods collectively enable running models up to twice the size of the available DRAM, with a 4-5x and 20-25x increase in inference speed compared to naive loading approaches in CPU and GPU, respectively. Our integration of sparsity awareness, context-adaptive loading, and a hardware-oriented design paves the way for effective inference of LLMs on devices with limited memory.

Large language model (LLM) scaling laws are empirical formulas that estimate changes in model quality as a result of increasing parameter count and training data. However, these formulas, including the popular DeepMind Chinchilla scaling laws, neglect to include the cost of inference. We modify the Chinchilla scaling laws to calculate the optimal LLM parameter count and pre-training data size to train and deploy a model of a given quality and inference demand. We conduct our analysis both in terms of a compute budget and real-world costs and find that LLM researchers expecting reasonably large inference demand (~1B requests) should train models smaller and longer than Chinchilla-optimal.

Deploying pre-trained transformer models like BERT on downstream tasks in resource-constrained scenarios is challenging due to their high inference cost, which grows rapidly with input sequence length. In this work, we propose a constraint-aware and ranking-distilled token pruning method ToP, which selectively removes unnecessary tokens as input sequence passes through layers, allowing the model to improve online inference speed while preserving accuracy. ToP overcomes the limitation of inaccurate token importance ranking in the conventional self-attention mechanism through a ranking-distilled token distillation technique, which distills effective token rankings from the final layer of unpruned models to early layers of pruned models. Then, ToP introduces a coarse-to-fine pruning approach that automatically selects the optimal subset of transformer layers and optimizes token pruning decisions within these layers through improved L_0 regularization. Extensive experiments on GLUE benchmark and SQuAD tasks demonstrate that ToP outperforms state-of-the-art token pruning and model compression methods with improved accuracy and speedups. ToP reduces the average FLOPs of BERT by 8.1x while achieving competitive accuracy on GLUE, and provides a real latency speedup of up to 7.4x on an Intel CPU.

Diffusion Transformers (DiTs) have gained prominence for outstanding scalability and extraordinary performance in generative tasks. However, their considerable inference costs impede practical deployment. The feature cache mechanism, which involves storing and retrieving redundant computations across timesteps, holds promise for reducing per-step inference time in diffusion models. Most existing caching methods for DiT are manually designed. Although the learning-based approach attempts to optimize strategies adaptively, it suffers from discrepancies between training and inference, which hampers both the performance and acceleration ratio. Upon detailed analysis, we pinpoint that these discrepancies primarily stem from two aspects: (1) Prior Timestep Disregard, where training ignores the effect of cache usage at earlier timesteps, and (2) Objective Mismatch, where the training target (align predicted noise in each timestep) deviates from the goal of inference (generate the high-quality image). To alleviate these discrepancies, we propose HarmoniCa, a novel method that Harmonizes training and inference with a novel learning-based Caching framework built upon Step-Wise Denoising Training (SDT) and Image Error Proxy-Guided Objective (IEPO). Compared to the traditional training paradigm, the newly proposed SDT maintains the continuity of the denoising process, enabling the model to leverage information from prior timesteps during training, similar to the way it operates during inference. Furthermore, we design IEPO, which integrates an efficient proxy mechanism to approximate the final image error caused by reusing the cached feature. Therefore, IEPO helps balance final image quality and cache utilization, resolving the issue of training that only considers the impact of cache usage on the predicted output at each timestep.

Large Language Models are growing in size, and we expect them to continue to do so, as larger models train quicker. However, this increase in size will severely impact inference costs. Therefore model compression is important, to retain the performance of larger models, but with a reduced cost of running them. In this thesis we explore the methods of model compression, and we empirically demonstrate that the simple method of skipping latter attention sublayers in Transformer LLMs is an effective method of model compression, as these layers prove to be redundant, whilst also being incredibly computationally expensive. We observed a 21% speed increase in one-token generation for Llama 2 7B, whilst surprisingly and unexpectedly improving performance over several common benchmarks.

Bayesian neural networks (BNNs) offer uncertainty quantification but come with the downside of substantially increased training and inference costs. Sparse BNNs have been investigated for efficient inference, typically by either slowly introducing sparsity throughout the training or by post-training compression of dense BNNs. The dilemma of how to cut down massive training costs remains, particularly given the requirement to learn about the uncertainty. To solve this challenge, we introduce Sparse Subspace Variational Inference (SSVI), the first fully sparse BNN framework that maintains a consistently highly sparse Bayesian model throughout the training and inference phases. Starting from a randomly initialized low-dimensional sparse subspace, our approach alternately optimizes the sparse subspace basis selection and its associated parameters. While basis selection is characterized as a non-differentiable problem, we approximate the optimal solution with a removal-and-addition strategy, guided by novel criteria based on weight distribution statistics. Our extensive experiments show that SSVI sets new benchmarks in crafting sparse BNNs, achieving, for instance, a 10-20x compression in model size with under 3\% performance drop, and up to 20x FLOPs reduction during training compared with dense VI training. Remarkably, SSVI also demonstrates enhanced robustness to hyperparameters, reducing the need for intricate tuning in VI and occasionally even surpassing VI-trained dense BNNs on both accuracy and uncertainty metrics.

Small, highly trained, open-source large language models are widely used due to their inference efficiency, but further improving their quality remains a challenge. Sparse upcycling is a promising approach that transforms a pretrained dense model into a Mixture-of-Experts (MoE) architecture, increasing the model's parameter count and quality. In this work, we compare the effectiveness of sparse upcycling against continued pretraining (CPT) across different model sizes, compute budgets, and pretraining durations. Our experiments show that sparse upcycling can achieve better quality, with improvements of over 20% relative to CPT in certain scenarios. However, this comes with a significant inference cost, leading to 40% slowdowns in high-demand inference settings for larger models. Our findings highlight the trade-off between model quality and inference efficiency, offering insights for practitioners seeking to balance model quality and deployment constraints.

Convolutional layers are one of the basic building blocks of modern deep neural networks. One fundamental assumption is that convolutional kernels should be shared for all examples in a dataset. We propose conditionally parameterized convolutions (CondConv), which learn specialized convolutional kernels for each example. Replacing normal convolutions with CondConv enables us to increase the size and capacity of a network, while maintaining efficient inference. We demonstrate that scaling networks with CondConv improves the performance and inference cost trade-off of several existing convolutional neural network architectures on both classification and detection tasks. On ImageNet classification, our CondConv approach applied to EfficientNet-B0 achieves state-of-the-art performance of 78.3% accuracy with only 413M multiply-adds. Code and checkpoints for the CondConv Tensorflow layer and CondConv-EfficientNet models are available at: https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/condconv.

Scaling laws are powerful tools to predict the performance of large language models. However, current scaling laws fall short of accounting for inference costs. In this work, we first show that model architecture affects inference latency, where models of the same size can have up to 3.5x difference in latency. To tackle this challenge, we modify the Chinchilla scaling laws to co-optimize the model parameter count, the number of training tokens, and the model architecture. Due to the reason that models of similar training loss exhibit gaps in downstream evaluation, we also propose a novel method to train inference-efficient models based on the revised scaling laws. We perform extensive empirical studies to fit and evaluate our inference-aware scaling laws. We vary model parameters from 80M to 1B, training tokens from 1.6B to 30B, and model shapes, training a total of 63 models. Guided by our inference-efficient scaling law and model selection method, we release the Morph-1B model, which improves inference latency by 1.8x while maintaining accuracy on downstream tasks compared to open-source models, pushing the Pareto frontier of accuracy-latency tradeoff.

Powerful foundation models, including large language models (LLMs), with Transformer architectures have ushered in a new era of Generative AI across various industries. Industry and research community have witnessed a large number of new applications, based on those foundation models. Such applications include question and answer, customer services, image and video generation, and code completions, among others. However, as the number of model parameters reaches to hundreds of billions, their deployment incurs prohibitive inference costs and high latency in real-world scenarios. As a result, the demand for cost-effective and fast inference using AI accelerators is ever more higher. To this end, our tutorial offers a comprehensive discussion on complementary inference optimization techniques using AI accelerators. Beginning with an overview of basic Transformer architectures and deep learning system frameworks, we deep dive into system optimization techniques for fast and memory-efficient attention computations and discuss how they can be implemented efficiently on AI accelerators. Next, we describe architectural elements that are key for fast transformer inference. Finally, we examine various model compression and fast decoding strategies in the same context.

The dominance of proprietary LLMs has led to restricted access and raised information privacy concerns. High-performing open-source alternatives are crucial for information-sensitive and high-volume applications but often lag behind in performance. To address this gap, we propose (1) A untargeted variant of iterative self-critique and self-refinement devoid of external influence. (2) A novel ranking metric - Performance, Refinement, and Inference Cost Score (PeRFICS) - to find the optimal model for a given task considering refined performance and cost. Our experiments show that SoTA open source models of varying sizes from 7B - 65B, on average, improve 8.2% from their baseline performance. Strikingly, even models with extremely small memory footprints, such as Vicuna-7B, show a 11.74% improvement overall and up to a 25.39% improvement in high-creativity, open ended tasks on the Vicuna benchmark. Vicuna-13B takes it a step further and outperforms ChatGPT post-refinement. This work has profound implications for resource-constrained and information-sensitive environments seeking to leverage LLMs without incurring prohibitive costs, compromising on performance and privacy. The domain-agnostic self-refinement process coupled with our novel ranking metric facilitates informed decision-making in model selection, thereby reducing costs and democratizing access to high-performing language models, as evidenced by case studies.

The sequential nature of modern LLMs makes them expensive and slow, and speculative sampling has proven to be an effective solution to this problem. Methods like EAGLE perform autoregression at the feature level, reusing top-layer features from the target model to achieve better results than vanilla speculative sampling. A growing trend in the LLM community is scaling up training data to improve model intelligence without increasing inference costs. However, we observe that scaling up data provides limited improvements for EAGLE. We identify that this limitation arises from EAGLE's feature prediction constraints. In this paper, we introduce EAGLE-3, which abandons feature prediction in favor of direct token prediction and replaces reliance on top-layer features with multi-layer feature fusion via a technique named training-time test. These improvements significantly enhance performance and enable the draft model to fully benefit from scaling up training data. Our experiments include both chat models and reasoning models, evaluated on five tasks. The results show that EAGLE-3 achieves a speedup ratio up to 6.5x, with about 1.4x improvement over EAGLE-2. The code is available at https://github.com/SafeAILab/EAGLE.

Vision-Language Models (VLMs) excel across diverse tasks but suffer from high inference costs in time and memory. Token sparsity mitigates inefficiencies in token usage, while neuron sparsity reduces high-dimensional computations, both offering promising solutions to enhance efficiency. Recently, these two sparsity paradigms have evolved largely in parallel, fostering the prevailing assumption that they function independently. However, a fundamental yet underexplored question remains: Do they truly operate in isolation, or is there a deeper underlying interplay that has yet to be uncovered? In this paper, we conduct the first comprehensive investigation into this question. By introducing and analyzing the matching mechanism between Core Neurons and Core Tokens, we found that key neurons and tokens for inference mutually influence and reinforce each other. Building on this insight, we propose CoreMatching, a co-adaptive sparse inference framework, which leverages the synergy between token and neuron sparsity to enhance inference efficiency. Through theoretical analysis and efficiency evaluations, we demonstrate that the proposed method surpasses state-of-the-art baselines on ten image understanding tasks and three hardware devices. Notably, on the NVIDIA Titan Xp, it achieved 5x FLOPs reduction and a 10x overall speedup. Code is released at https://github.com/wangqinsi1/2025-ICML-CoreMatching/tree/main.

Performing inference on hundreds of thousands of samples with large language models (LLMs) can be computationally and financially costly. We propose batch prompting, a simple alternative prompting approach that enables the LLM to run inference in batches, instead of one sample at a time. Our method reduces both token and time costs while retaining downstream performance. We theoretically demonstrate that under a few-shot in-context learning setting, the inference costs decrease almost inverse linearly with the number of samples in each batch. We extensively validate the effectiveness of batch prompting on ten datasets across commonsense QA, arithmetic reasoning, and NLI/NLU: batch prompting significantly~(up to 5times with six samples in batch) reduces the LLM (Codex) inference token and time costs while achieving better or comparable performance. Our analysis shows that the number of samples in each batch and the complexity of tasks affect its performance. Further, batch prompting can be applied across different LLMs and reasoning methods.

Large language models (LLMs) are widely applied in chatbots, code generators, and search engines. Workloads such as chain-of-thought, complex reasoning, and agent services significantly increase the inference cost by invoking the model repeatedly. Optimization methods such as parallelism, compression, and caching have been adopted to reduce costs, but the diverse service requirements make it hard to select the right method. Recently, specialized LLM inference engines have emerged as a key component for integrating the optimization methods into service-oriented infrastructures. However, a systematic study on inference engines is still lacking. This paper provides a comprehensive evaluation of 25 open-source and commercial inference engines. We examine each inference engine in terms of ease-of-use, ease-of-deployment, general-purpose support, scalability, and suitability for throughput- and latency-aware computation. Furthermore, we explore the design goals of each inference engine by investigating the optimization techniques it supports. In addition, we assess the ecosystem maturity of open source inference engines and handle the performance and cost policy of commercial solutions. We outline future research directions that include support for complex LLM-based services, support of various hardware, and enhanced security, offering practical guidance to researchers and developers in selecting and designing optimized LLM inference engines. We also provide a public repository to continually track developments in this fast-evolving field: https://github.com/sihyeong/Awesome-LLM-Inference-Engine

Recent years have seen a surge in the popularity of commercial AI products based on generative, multi-purpose AI systems promising a unified approach to building machine learning (ML) models into technology. However, this ambition of "generality" comes at a steep cost to the environment, given the amount of energy these systems require and the amount of carbon that they emit. In this work, we propose the first systematic comparison of the ongoing inference cost of various categories of ML systems, covering both task-specific (i.e. finetuned models that carry out a single task) and `general-purpose' models, (i.e. those trained for multiple tasks). We measure deployment cost as the amount of energy and carbon required to perform 1,000 inferences on representative benchmark dataset using these models. We find that multi-purpose, generative architectures are orders of magnitude more expensive than task-specific systems for a variety of tasks, even when controlling for the number of model parameters. We conclude with a discussion around the current trend of deploying multi-purpose generative ML systems, and caution that their utility should be more intentionally weighed against increased costs in terms of energy and emissions. All the data from our study can be accessed via an interactive demo to carry out further exploration and analysis.

Diffusion models have achieved impressive advancements in various vision tasks. However, these gains often rely on increasing model size, which escalates computational complexity and memory demands, complicating deployment, raising inference costs, and causing environmental impact. While some studies have explored pruning techniques to improve the memory efficiency of diffusion models, most existing methods require extensive retraining to retain the model performance. Retraining a modern large diffusion model is extremely costly and resource-intensive, which limits the practicality of these methods. In this work, we achieve low-cost diffusion pruning without retraining by proposing a model-agnostic structural pruning framework for diffusion models that learns a differentiable mask to sparsify the model. To ensure effective pruning that preserves the quality of the final denoised latent, we design a novel end-to-end pruning objective that spans the entire diffusion process. As end-to-end pruning is memory-intensive, we further propose time step gradient checkpointing, a technique that significantly reduces memory usage during optimization, enabling end-to-end pruning within a limited memory budget. Results on state-of-the-art U-Net diffusion models SDXL and diffusion transformers (FLUX) demonstrate that our method can effectively prune up to 20% parameters with minimal perceptible performance degradation, and notably, without the need for model retraining. We also showcase that our method can still prune on top of time step distilled diffusion models.

Fine-tuning large models is highly effective, however, inference can be expensive and produces carbon emissions. Knowledge distillation has been shown to be a practical solution to reduce inference costs, but the distillation process itself requires significant computational resources. Rather than buying or renting GPUs to fine-tune, then distill a large model, an NLP practitioner might instead choose to allocate the available budget to hire annotators and manually label additional fine-tuning data. In this paper, we investigate how to most efficiently use a fixed budget to build a compact model. Through extensive experiments on six diverse tasks, we show that distilling from T5-XXL (11B) to T5-Small (60M) is almost always a cost-efficient strategy compared to annotating more data to directly train a compact model (T5-Small). We further investigate how the optimal budget allocated towards computation varies across scenarios. We will make our code, datasets, annotation cost estimates, and baseline models available as a benchmark to support further work on cost-efficient training of compact models.

The attention mechanism of a transformer has a quadratic complexity, leading to high inference costs and latency for long sequences. However, attention matrices are mostly sparse, which implies that many entries may be omitted from computation for efficient inference. Sparse attention inference methods aim to reduce this computational burden; however, they also come with a troublesome performance degradation. We discover that one reason for this degradation is that the sparse calculation induces a distributional shift in the attention outputs. The distributional shift causes decoding-time queries to fail to align well with the appropriate keys from the prefill stage, leading to a drop in performance. We propose a simple, novel, and effective procedure for correcting this distributional shift, bringing the distribution of sparse attention outputs closer to that of quadratic attention. Our method can be applied on top of any sparse attention method, and results in an average 36%pt performance increase, recovering 88% of quadratic attention accuracy on the 131K RULER benchmark when applied on top of sliding window attention with sink tokens while only adding a small overhead. Our method can maintain approximately 98.5% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.

Scaling test-time compute has emerged as a key ingredient for enabling large language models (LLMs) to solve difficult problems, but comes with high latency and inference cost. We introduce sleep-time compute, which allows models to "think" offline about contexts before queries are presented: by anticipating what queries users might ask and pre-computing useful quantities, we can significantly reduce the compute requirements at test-time. To demonstrate the efficacy of our method, we create modified versions of two reasoning tasks - Stateful GSM-Symbolic and Stateful AIME. We find that sleep-time compute can reduce the amount of test-time compute needed to achieve the same accuracy by ~ 5x on Stateful GSM-Symbolic and Stateful AIME and that by scaling sleep-time compute we can further increase accuracy by up to 13% on Stateful GSM-Symbolic and 18% on Stateful AIME. Furthermore, we introduce Multi-Query GSM-Symbolic, which extends GSM-Symbolic by including multiple related queries per context. By amortizing sleep-time compute across related queries about the same context using Multi-Query GSM-Symbolic, we can decrease the average cost per query by 2.5x. We then conduct additional analysis to understand when sleep-time compute is most effective, finding the predictability of the user query to be well correlated with the efficacy of sleep-time compute. Finally, we conduct a case-study of applying sleep-time compute to a realistic agentic SWE task.

Multi-head Latent Attention (MLA) is an innovative architecture proposed by DeepSeek, designed to ensure efficient and economical inference by significantly compressing the Key-Value (KV) cache into a latent vector. Compared to MLA, standard LLMs employing Multi-Head Attention (MHA) and its variants such as Grouped-Query Attention (GQA) exhibit significant cost disadvantages. Enabling well-trained LLMs (e.g., Llama) to rapidly adapt to MLA without pre-training from scratch is both meaningful and challenging. This paper proposes the first data-efficient fine-tuning method for transitioning from MHA to MLA (MHA2MLA), which includes two key components: for partial-RoPE, we remove RoPE from dimensions of queries and keys that contribute less to the attention scores, for low-rank approximation, we introduce joint SVD approximations based on the pre-trained parameters of keys and values. These carefully designed strategies enable MHA2MLA to recover performance using only a small fraction (0.3% to 0.6%) of the data, significantly reducing inference costs while seamlessly integrating with compression techniques such as KV cache quantization. For example, the KV cache size of Llama2-7B is reduced by 92.19%, with only a 0.5% drop in LongBench performance.

Leveraging the diffusion transformer (DiT) architecture, models like Sora, CogVideoX and Wan have achieved remarkable progress in text-to-video, image-to-video, and video editing tasks. Despite these advances, diffusion-based video generation remains computationally intensive, especially for high-resolution, long-duration videos. Prior work accelerates its inference by skipping computation, usually at the cost of severe quality degradation. In this paper, we propose SRDiffusion, a novel framework that leverages collaboration between large and small models to reduce inference cost. The large model handles high-noise steps to ensure semantic and motion fidelity (Sketching), while the smaller model refines visual details in low-noise steps (Rendering). Experimental results demonstrate that our method outperforms existing approaches, over 3times speedup for Wan with nearly no quality loss for VBench, and 2times speedup for CogVideoX. Our method is introduced as a new direction orthogonal to existing acceleration strategies, offering a practical solution for scalable video generation.

Recently, large-scale transformer-based models have been proven to be effective over various tasks across many domains. Nevertheless, applying them in industrial production requires tedious and heavy works to reduce inference costs. To fill such a gap, we introduce a scalable inference solution: Easy and Efficient Transformer (EET), including a series of transformer inference optimization at the algorithm and implementation levels. First, we design highly optimized kernels for long inputs and large hidden sizes. Second, we propose a flexible CUDA memory manager to reduce the memory footprint when deploying a large model. Compared with the state-of-the-art transformer inference library (Faster Transformer v4.0), EET can achieve an average of 1.40-4.20x speedup on the transformer decoder layer with an A100 GPU

Large Language Models (LLMs) and other large foundation models have achieved noteworthy success, but their size exacerbates existing resource consumption and latency challenges. In particular, the large-scale deployment of these models is hindered by the significant resource requirements during inference. In this paper, we study two approaches for mitigating these challenges: employing a cache to store previous queries and learning a model multiplexer to choose from an ensemble of models for query processing. Theoretically, we provide an optimal algorithm for jointly optimizing both approaches to reduce the inference cost in both offline and online tabular settings. By combining a caching algorithm, namely Greedy Dual Size with Frequency (GDSF) or Least Expected Cost (LEC), with a model multiplexer, we achieve optimal rates in both offline and online settings. Empirically, simulations show that the combination of our caching and model multiplexing algorithms greatly improves over the baselines, with up to 50times improvement over the baseline when the ratio between the maximum cost and minimum cost is 100. Experiments on real datasets show a 4.3times improvement in FLOPs over the baseline when the ratio for FLOPs is 10, and a 1.8times improvement in latency when the ratio for average latency is 1.85.

Specializing LLMs in various domain-specific tasks has emerged as a critical step towards achieving high performance. However, the construction and annotation of datasets in specific domains are always very costly. Apart from using superior and expensive closed-source LLM APIs to construct datasets, some open-source models have become strong enough to handle dataset construction in many scenarios. Thus, we present a family of data augmentation models designed to significantly improve the efficiency for model fine-tuning. These models, trained based on sufficiently small LLMs, support key functionalities with low inference costs: instruction expansion, instruction refinement, and instruction-response pair expansion. To fulfill this goal, we first construct an automatic data collection system with seed datasets generated from both public repositories and our in-house datasets. This system leverages powerful LLMs to expand, refine and re-write the instructions and responses, incorporating quality assessment techniques. Following this, we introduce the training process of our models, which effectively distills task-solving and text synthesis abilities from teacher LLMs. Finally, we demonstrate how we integrate these functionalities into a machine learning platform to support low-cost LLM fine-tuning from both dataset preparation and training perspectives for users. Experiments and an application study prove the effectiveness of our approach.

Emerging Large Language Model (LLM) applications require long input prompts to perform complex downstream tasks like document analysis and code generation. For these long context length applications, the length of the input prompt poses a significant challenge in terms of inference efficiency since the inference costs increase linearly with sequence length. However, for many of these applications, much of the context in the prompt is fixed across different user inputs, thereby providing the opportunity to perform offline optimizations to process user inputs quickly, as they are received. In this work, we propose Squeezed Attention as a mechanism to accelerate LLM applications where a large portion of the input prompt is fixed. We first leverage K-means clustering offline to group the keys for the fixed context based on semantic similarity and represent each cluster with a single centroid value. During inference, we compare query tokens from the user input with the centroids to predict which of the keys from the fixed context are semantically relevant and need to be loaded during inference. We then compute exact attention using only these important keys from the fixed context, thereby reducing bandwidth and computational costs. We also extend our method to use a hierarchical centroid lookup to identify important keys, which can reduce the complexity of attention from linear to logarithmic with respect to the context length. We implement optimized Triton kernels for centroid comparison and sparse FlashAttention with important keys, achieving more than 4x speedups during both the prefill and generation phases for long-context inference. Furthermore, we have extensively evaluated our method on various long-context benchmarks including LongBench, where it achieves a 3x reduction in KV cache budget without accuracy loss and up to an 8x reduction with <0.5 point accuracy gap for various models.

The computational cost of transformer models makes them inefficient in low-latency or low-power applications. While techniques such as quantization or linear attention can reduce the computational load, they may incur a reduction in accuracy. In addition, globally reducing the cost for all inputs may be sub-optimal. We observe that for each layer, the full width of the layer may be needed only for a small subset of tokens inside a batch and that the "effective" width needed to process a token can vary from layer to layer. Motivated by this observation, we introduce the Adaptive Computation Module (ACM), a generic module that dynamically adapts its computational load to match the estimated difficulty of the input on a per-token basis. An ACM consists of a sequence of learners that progressively refine the output of their preceding counterparts. An additional gating mechanism determines the optimal number of learners to execute for each token. We also describe a distillation technique to replace any pre-trained model with an "ACMized" variant. The distillation phase is designed to be highly parallelizable across layers while being simple to plug-and-play into existing networks. Our evaluation of transformer models in computer vision and speech recognition demonstrates that substituting layers with ACMs significantly reduces inference costs without degrading the downstream accuracy for a wide interval of user-defined budgets.

The recent advancements in large language models (LLMs) have been extraordinary, yet the escalating inference costs associated with them present challenges in real-world applications. To address these challenges, we propose a novel approach called Early-exiting Speculative Decoding (EESD) with lossless acceleration. Specifically, EESD utilizes a segment of the LLM to generate draft tokens, incorporating Early-exiting structures after the first N layers. To enhance the quality of draft tokens, a self-distillation method is integrated. This early-exiting design not only reduces deployment and training costs but also significantly accelerates the token generation speed. Moreover, we introduce a novel sampling mechanism that leverages Thompson Sampling to regulate the generation processes, automatically determining the quantity of draft tokens in each round. The original LLM is then employed to validate these draft tokens through a single forward pass, and thus guarantees that the final output text maintains a distribution consistent with vanilla auto-regressive decoding. The experimental results on both 13B and 70B models demonstrate that our approach decodes tokens at a markedly accelerated rate compared to prior methods, showing the effectiveness of our approach.

Multimodal models have gained significant success in recent years. Standard multimodal approaches often assume unchanged modalities from training stage to inference stage. In practice, however, many scenarios fail to satisfy such assumptions with missing modalities during inference, leading to limitations on where multimodal models can be applied. While existing methods mitigate the problem through reconstructing the missing modalities, it increases unnecessary computational cost, which could be just as critical, especially for large, deployed systems. To solve the problem from both sides, we propose a novel guidance network that promotes knowledge sharing during training, taking advantage of the multimodal representations to train better single-modality models for inference. Real-life experiment in violence detection shows that our proposed framework trains single-modality models that significantly outperform its traditionally trained counterparts while maintaining the same inference cost.

There is a rapidly growing number of large language models (LLMs) that users can query for a fee. We review the cost associated with querying popular LLM APIs, e.g. GPT-4, ChatGPT, J1-Jumbo, and find that these models have heterogeneous pricing structures, with fees that can differ by two orders of magnitude. In particular, using LLMs on large collections of queries and text can be expensive. Motivated by this, we outline and discuss three types of strategies that users can exploit to reduce the inference cost associated with using LLMs: 1) prompt adaptation, 2) LLM approximation, and 3) LLM cascade. As an example, we propose FrugalGPT, a simple yet flexible instantiation of LLM cascade which learns which combinations of LLMs to use for different queries in order to reduce cost and improve accuracy. Our experiments show that FrugalGPT can match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction or improve the accuracy over GPT-4 by 4% with the same cost. The ideas and findings presented here lay a foundation for using LLMs sustainably and efficiently.

Sparse Mixture-of-Experts (MoE) has been a successful approach for scaling multilingual translation models to billions of parameters without a proportional increase in training computation. However, MoE models are prohibitively large and practitioners often resort to methods such as distillation for serving. In this work, we investigate routing strategies at different granularity (token, sentence, task) in MoE models to bypass distillation. Experiments on WMT and a web-scale dataset suggest that task-level routing (task-MoE) enables us to extract smaller, ready-to-deploy sub-networks from large sparse models. On WMT, our task-MoE with 32 experts (533M parameters) outperforms the best performing token-level MoE model (token-MoE) by +1.0 BLEU on average across 30 language pairs. The peak inference throughput is also improved by a factor of 1.9x when we route by tasks instead of tokens. While distilling a token-MoE to a smaller dense model preserves only 32% of the BLEU gains, our sub-network task-MoE, by design, preserves all the gains with the same inference cost as the distilled student model. Finally, when scaling up to 200 language pairs, our 128-expert task-MoE (13B parameters) performs competitively with a token-level counterpart, while improving the peak inference throughput by a factor of 2.6x.

The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we may average many models without incurring any additional inference or memory costs -- we call the results "model soups." When fine-tuning large pre-trained models such as CLIP, ALIGN, and a ViT-G pre-trained on JFT, our soup recipe provides significant improvements over the best model in a hyperparameter sweep on ImageNet. The resulting ViT-G model, which attains 90.94% top-1 accuracy on ImageNet, achieved a new state of the art. Furthermore, we show that the model soup approach extends to multiple image classification and natural language processing tasks, improves out-of-distribution performance, and improves zero-shot performance on new downstream tasks. Finally, we analytically relate the performance similarity of weight-averaging and logit-ensembling to flatness of the loss and confidence of the predictions, and validate this relation empirically. Code is available at https://github.com/mlfoundations/model-soups.

Despite the successes of large language models (LLMs), they exhibit significant drawbacks, particularly when processing long contexts. Their inference cost scales quadratically with respect to sequence length, making it expensive for deployment in some real-world text processing applications, such as retrieval-augmented generation (RAG). Additionally, LLMs also exhibit the "distraction phenomenon," where irrelevant context in the prompt degrades output quality. To address these drawbacks, we propose a novel RAG prompting methodology, superposition prompting, which can be directly applied to pre-trained transformer-based LLMs without the need for fine-tuning. At a high level, superposition prompting allows the LLM to process input documents in parallel prompt paths, discarding paths once they are deemed irrelevant. We demonstrate the capability of our method to simultaneously enhance time efficiency across a variety of question-answering benchmarks using multiple pre-trained LLMs. Furthermore, our technique significantly improves accuracy when the retrieved context is large relative the context the model was trained on. For example, our approach facilitates an 93x reduction in compute time while improving accuracy by 43\% on the NaturalQuestions-Open dataset with the MPT-7B instruction-tuned model over naive RAG.

Scaling model size and training data has led to great advances in the performance of Large Language Models (LLMs). However, the diminishing returns of this approach necessitate alternative methods to improve model capabilities, particularly in tasks requiring advanced reasoning. Large reasoning models, which leverage long chain-of-thoughts, bring unprecedented breakthroughs in problem-solving capabilities but at a substantial deployment cost associated to longer generations. Reducing inference costs is crucial for the economic feasibility, user experience, and environmental sustainability of these models. In this work, we propose to train large reasoning models to reason efficiently. More precisely, we use reinforcement learning (RL) to train reasoning models to dynamically allocate inference-time compute based on task complexity. Our method incentivizes models to minimize unnecessary computational overhead while maintaining accuracy, thereby achieving substantial efficiency gains. It enables the derivation of a family of reasoning models with varying efficiency levels, controlled via a single hyperparameter. Experiments on two open-weight large reasoning models demonstrate significant reductions in inference cost while preserving most of the accuracy.

We explore an evolutionary search strategy for scaling inference time compute in Large Language Models. The proposed approach, Mind Evolution, uses a language model to generate, recombine and refine candidate responses. The proposed approach avoids the need to formalize the underlying inference problem whenever a solution evaluator is available. Controlling for inference cost, we find that Mind Evolution significantly outperforms other inference strategies such as Best-of-N and Sequential Revision in natural language planning tasks. In the TravelPlanner and Natural Plan benchmarks, Mind Evolution solves more than 98% of the problem instances using Gemini 1.5 Pro without the use of a formal solver.

Few-shot dialogue state tracking (DST) with Large Language Models (LLM) relies on an effective and efficient conversation retriever to find similar in-context examples for prompt learning. Previous works use raw dialogue context as search keys and queries, and a retriever is fine-tuned with annotated dialogues to achieve superior performance. However, the approach is less suited for scaling to new domains or new annotation languages, where fine-tuning data is unavailable. To address this problem, we handle the task of conversation retrieval based on text summaries of the conversations. A LLM-based conversation summarizer is adopted for query and key generation, which enables effective maximum inner product search. To avoid the extra inference cost brought by LLM-based conversation summarization, we further distill a light-weight conversation encoder which produces query embeddings without decoding summaries for test conversations. We validate our retrieval approach on MultiWOZ datasets with GPT-Neo-2.7B and LLaMA-7B/30B. The experimental results show a significant improvement over relevant baselines in real few-shot DST settings.

Vision Transformers (ViT) have emerged as the de-facto choice for numerous industry grade vision solutions. But their inference cost can be prohibitive for many settings, as they compute self-attention in each layer which suffers from quadratic computational complexity in the number of tokens. On the other hand, spatial information in images and spatio-temporal information in videos is usually sparse and redundant. In this work, we introduce LookupViT, that aims to exploit this information sparsity to reduce ViT inference cost. LookupViT provides a novel general purpose vision transformer block that operates by compressing information from higher resolution tokens to a fixed number of tokens. These few compressed tokens undergo meticulous processing, while the higher-resolution tokens are passed through computationally cheaper layers. Information sharing between these two token sets is enabled through a bidirectional cross-attention mechanism. The approach offers multiple advantages - (a) easy to implement on standard ML accelerators (GPUs/TPUs) via standard high-level operators, (b) applicable to standard ViT and its variants, thus generalizes to various tasks, (c) can handle different tokenization and attention approaches. LookupViT also offers flexibility for the compressed tokens, enabling performance-computation trade-offs in a single trained model. We show LookupViT's effectiveness on multiple domains - (a) for image-classification (ImageNet-1K and ImageNet-21K), (b) video classification (Kinetics400 and Something-Something V2), (c) image captioning (COCO-Captions) with a frozen encoder. LookupViT provides 2times reduction in FLOPs while upholding or improving accuracy across these domains. In addition, LookupViT also demonstrates out-of-the-box robustness and generalization on image classification (ImageNet-C,R,A,O), improving by up to 4% over ViT.

To reduce the inference cost of large language models, model compression is increasingly used to create smaller scalable models. However, little is known about their robustness to minority subgroups defined by the labels and attributes of a dataset. In this paper, we investigate the effects of 18 different compression methods and settings on the subgroup robustness of BERT language models. We show that worst-group performance does not depend on model size alone, but also on the compression method used. Additionally, we find that model compression does not always worsen the performance on minority subgroups. Altogether, our analysis serves to further research into the subgroup robustness of model compression.

The self-attention mechanism in the transformer architecture is capable of capturing long-range dependencies and it is the main reason behind its effectiveness in processing sequential data. Nevertheless, despite their success, transformers have two significant drawbacks that still limit their broader applicability: (1) In order to remember past information, the self-attention mechanism requires access to the whole history to be provided as context. (2) The inference cost in transformers is expensive. In this paper we introduce recurrent alternatives to the transformer self-attention mechanism that offer a context-independent inference cost, leverage long-range dependencies effectively, and perform well in practice. We evaluate our approaches in reinforcement learning problems where the aforementioned computational limitations make the application of transformers nearly infeasible. We quantify the impact of the different components of our architecture in a diagnostic environment and assess performance gains in 2D and 3D pixel-based partially-observable environments. When compared to a state-of-the-art architecture, GTrXL, inference in our approach is at least 40% cheaper while reducing memory use in more than 50%. Our approach either performs similarly or better than GTrXL, improving more than 37% upon GTrXL performance on harder tasks.

We introduce the Byte Latent Transformer (BLT), a new byte-level LLM architecture that, for the first time, matches tokenization-based LLM performance at scale with significant improvements in inference efficiency and robustness. BLT encodes bytes into dynamically sized patches, which serve as the primary units of computation. Patches are segmented based on the entropy of the next byte, allocating more compute and model capacity where increased data complexity demands it. We present the first FLOP controlled scaling study of byte-level models up to 8B parameters and 4T training bytes. Our results demonstrate the feasibility of scaling models trained on raw bytes without a fixed vocabulary. Both training and inference efficiency improve due to dynamically selecting long patches when data is predictable, along with qualitative improvements on reasoning and long tail generalization. Overall, for fixed inference costs, BLT shows significantly better scaling than tokenization-based models, by simultaneously growing both patch and model size.

We introduce Qwen2.5-1M, a series of models that extend the context length to 1 million tokens. Compared to the previous 128K version, the Qwen2.5-1M series have significantly enhanced long-context capabilities through long-context pre-training and post-training. Key techniques such as long data synthesis, progressive pre-training, and multi-stage supervised fine-tuning are employed to effectively enhance long-context performance while reducing training costs. To promote the use of long-context models among a broader user base, we present and open-source our inference framework. This framework includes a length extrapolation method that can expand the model context lengths by at least four times, or even more, without additional training. To reduce inference costs, we implement a sparse attention method along with chunked prefill optimization for deployment scenarios and a sparsity refinement method to improve precision. Additionally, we detail our optimizations in the inference engine, including kernel optimization, pipeline parallelism, and scheduling optimization, which significantly enhance overall inference performance. By leveraging our inference framework, the Qwen2.5-1M models achieve a remarkable 3x to 7x prefill speedup in scenarios with 1 million tokens of context. This framework provides an efficient and powerful solution for developing applications that require long-context processing using open-source models. The Qwen2.5-1M series currently includes the open-source models Qwen2.5-7B-Instruct-1M and Qwen2.5-14B-Instruct-1M, as well as the API-accessed model Qwen2.5-Turbo. Evaluations show that Qwen2.5-1M models have been greatly improved in long-context tasks without compromising performance in short-context scenarios. Specifically, the Qwen2.5-14B-Instruct-1M model significantly outperforms GPT-4o-mini in long-context tasks and supports contexts eight times longer.

Large language models are well-known to be effective at few-shot in-context learning (ICL). Recent advancements in multimodal foundation models have enabled unprecedentedly long context windows, presenting an opportunity to explore their capability to perform ICL with many more demonstrating examples. In this work, we evaluate the performance of multimodal foundation models scaling from few-shot to many-shot ICL. We benchmark GPT-4o and Gemini 1.5 Pro across 10 datasets spanning multiple domains (natural imagery, medical imagery, remote sensing, and molecular imagery) and tasks (multi-class, multi-label, and fine-grained classification). We observe that many-shot ICL, including up to almost 2,000 multimodal demonstrating examples, leads to substantial improvements compared to few-shot (<100 examples) ICL across all of the datasets. Further, Gemini 1.5 Pro performance continues to improve log-linearly up to the maximum number of tested examples on many datasets. Given the high inference costs associated with the long prompts required for many-shot ICL, we also explore the impact of batching multiple queries in a single API call. We show that batching up to 50 queries can lead to performance improvements under zero-shot and many-shot ICL, with substantial gains in the zero-shot setting on multiple datasets, while drastically reducing per-query cost and latency. Finally, we measure ICL data efficiency of the models, or the rate at which the models learn from more demonstrating examples. We find that while GPT-4o and Gemini 1.5 Pro achieve similar zero-shot performance across the datasets, Gemini 1.5 Pro exhibits higher ICL data efficiency than GPT-4o on most datasets. Our results suggest that many-shot ICL could enable users to efficiently adapt multimodal foundation models to new applications and domains. Our codebase is publicly available at https://github.com/stanfordmlgroup/ManyICL .

Visual token reduction lowers inference costs caused by extensive image features in large vision-language models (LVLMs). Unlike relevant studies that prune tokens in self-attention-only LVLMs, our work uniquely addresses cross-attention-based models, which achieve superior performance. We identify that the key-value (KV) cache size for image tokens in cross-attention layers significantly exceeds that of text tokens in self-attention layers, posing a major compute bottleneck. To mitigate this issue, we exploit the sparse nature in cross-attention maps to selectively prune redundant visual features. Our Trimmed Llama effectively reduces KV cache demands without requiring additional training. By benefiting from 50%-reduced visual features, our model can reduce inference latency and memory usage while achieving benchmark parity.

Human matting is a foundation task in image and video processing, where human foreground pixels are extracted from the input. Prior works either improve the accuracy by additional guidance or improve the temporal consistency of a single instance across frames. We propose a new framework MaGGIe, Masked Guided Gradual Human Instance Matting, which predicts alpha mattes progressively for each human instances while maintaining the computational cost, precision, and consistency. Our method leverages modern architectures, including transformer attention and sparse convolution, to output all instance mattes simultaneously without exploding memory and latency. Although keeping constant inference costs in the multiple-instance scenario, our framework achieves robust and versatile performance on our proposed synthesized benchmarks. With the higher quality image and video matting benchmarks, the novel multi-instance synthesis approach from publicly available sources is introduced to increase the generalization of models in real-world scenarios.

Layer fusion techniques are critical to improving the inference efficiency of deep neural networks (DNN) for deployment. Fusion aims to lower inference costs by reducing data transactions between an accelerator's on-chip buffer and DRAM. This is accomplished by grouped execution of multiple operations like convolution and activations together into single execution units - fusion groups. However, on-chip buffer capacity limits fusion group size and optimizing fusion on whole DNNs requires partitioning into multiple fusion groups. Finding the optimal groups is a complex problem where the presence of invalid solutions hampers traditional search algorithms and demands robust approaches. In this paper we incorporate Explainable AI, specifically Graph Explanation Techniques (GET), into layer fusion. Given an invalid fusion group, we identify the operations most responsible for group invalidity, then use this knowledge to recursively split the original fusion group via a greedy tree-based algorithm to minimize DRAM access. We pair our scheme with common algorithms and optimize DNNs on two types of layer fusion: Line-Buffer Depth First (LBDF) and Branch Requirement Reduction (BRR). Experiments demonstrate the efficacy of our scheme on several popular and classical convolutional neural networks like ResNets and MobileNets. Our scheme achieves over 20% DRAM Access reduction on EfficientNet-B3.

Model efficiency is a critical aspect of developing and deploying machine learning models. Inference time and latency directly affect the user experience, and some applications have hard requirements. In addition to inference costs, model training also have direct financial and environmental impacts. Although there are numerous well-established metrics (cost indicators) for measuring model efficiency, researchers and practitioners often assume that these metrics are correlated with each other and report only few of them. In this paper, we thoroughly discuss common cost indicators, their advantages and disadvantages, and how they can contradict each other. We demonstrate how incomplete reporting of cost indicators can lead to partial conclusions and a blurred or incomplete picture of the practical considerations of different models. We further present suggestions to improve reporting of efficiency metrics.

Large Transformer models have been central to recent advances in natural language processing. The training and inference costs of these models, however, have grown rapidly and become prohibitively expensive. Here we aim to reduce the costs of Transformers by searching for a more efficient variant. Compared to previous approaches, our search is performed at a lower level, over the primitives that define a Transformer TensorFlow program. We identify an architecture, named Primer, that has a smaller training cost than the original Transformer and other variants for auto-regressive language modeling. Primer's improvements can be mostly attributed to two simple modifications: squaring ReLU activations and adding a depthwise convolution layer after each Q, K, and V projection in self-attention. Experiments show Primer's gains over Transformer increase as compute scale grows and follow a power law with respect to quality at optimal model sizes. We also verify empirically that Primer can be dropped into different codebases to significantly speed up training without additional tuning. For example, at a 500M parameter size, Primer improves the original T5 architecture on C4 auto-regressive language modeling, reducing the training cost by 4X. Furthermore, the reduced training cost means Primer needs much less compute to reach a target one-shot performance. For instance, in a 1.9B parameter configuration similar to GPT-3 XL, Primer uses 1/3 of the training compute to achieve the same one-shot performance as Transformer. We open source our models and several comparisons in T5 to help with reproducibility.

We introduce SigLIP 2, a family of new multilingual vision-language encoders that build on the success of the original SigLIP. In this second iteration, we extend the original image-text training objective with several prior, independently developed techniques into a unified recipe -- this includes captioning-based pretraining, self-supervised losses (self-distillation, masked prediction) and online data curation. With these changes, SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, including zero-shot classification, image-text retrieval, and transfer performance when extracting visual representations for Vision-Language Models (VLMs). Furthermore, the new training recipe leads to significant improvements on localization and dense prediction tasks. We also train variants which support multiple resolutions and preserve the input's native aspect ratio. Finally, we train on a more diverse data-mixture that includes de-biasing techniques, leading to much better multilingual understanding and improved fairness. To allow users to trade off inference cost with performance, we release model checkpoints at four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B).

Vision-Language Models (VLMs) such as CLIP are trained on large amounts of image-text pairs, resulting in remarkable generalization across several data distributions. The prohibitively expensive training and data collection/curation costs of these models make them valuable Intellectual Property (IP) for organizations. This motivates a vendor-client paradigm, where a vendor trains a large-scale VLM and grants only input-output access to clients on a pay-per-query basis in a black-box setting. The client aims to minimize inference cost by distilling the VLM to a student model using the limited available task-specific data, and further deploying this student model in the downstream application. While naive distillation largely improves the In-Domain (ID) accuracy of the student, it fails to transfer the superior out-of-distribution (OOD) generalization of the VLM teacher using the limited available labeled images. To mitigate this, we propose Vision-Language to Vision-Align, Distill, Predict (VL2V-ADiP), which first aligns the vision and language modalities of the teacher model with the vision modality of a pre-trained student model, and further distills the aligned VLM embeddings to the student. This maximally retains the pre-trained features of the student, while also incorporating the rich representations of the VLM image encoder and the superior generalization of the text embeddings. The proposed approach achieves state-of-the-art results on the standard Domain Generalization benchmarks in a black-box teacher setting, and also when weights of the VLM are accessible.

Predominant LLMs focus on high-resource languages while leaving low-resource languages, particularly those in Southeast Asia (SEA), underrepresented. In addition, those models are general-purpose and pay limited attention to the e-commerce domain. To overcome these limitations, we introduce Compass-v2, a lightweight Mixture-of-Experts (MoE) model specifically designed for Southeast Asian languages and e-commerce applications. To balance model performance and inference cost, the model is designed with 30B total parameters and 5B active parameters, incorporating both fine-grained and shared expert modules. To enhance multilingual performance, we curated and constructed a high-quality, industry-leading SEA dataset, to the best of our knowledge. To boost performance in the e-commerce domain, we built a dataset comprising hundreds of billions of tokens, sourced through external data mining and internal platform collection. Besides, we pioneered a hybrid reasoning model that supports both fast thinking and deep thinking within a unified framework to enhance the reasoning capabilities, diverging from the conventional industry practice of deploying two separate models. Through extensive experimental evaluations, our model demonstrates state-of-the-art SEA multilingual and e-commerce performance among sub-30B models, while maintaining significantly lower inference cost.

Auto-regressive large language models (LLMs) have yielded impressive performance in many real-world tasks. However, the new paradigm of these LLMs also exposes novel threats. In this paper, we explore their vulnerability to inference cost attacks, where a malicious user crafts Engorgio prompts to intentionally increase the computation cost and latency of the inference process. We design Engorgio, a novel methodology, to efficiently generate adversarial Engorgio prompts to affect the target LLM's service availability. Engorgio has the following two technical contributions. (1) We employ a parameterized distribution to track LLMs' prediction trajectory. (2) Targeting the auto-regressive nature of LLMs' inference process, we propose novel loss functions to stably suppress the appearance of the <EOS> token, whose occurrence will interrupt the LLM's generation process. We conduct extensive experiments on 13 open-sourced LLMs with parameters ranging from 125M to 30B. The results show that Engorgio prompts can successfully induce LLMs to generate abnormally long outputs (i.e., roughly 2-13times longer to reach 90%+ of the output length limit) in a white-box scenario and our real-world experiment demonstrates Engergio's threat to LLM service with limited computing resources. The code is accessible at https://github.com/jianshuod/Engorgio-prompt.

Large Language Models (LLMs) have achieved great success in solving difficult tasks across many domains, but such success comes with a high computation cost, and inference latency. As developers and third parties customize these models, the need to provide efficient inference has increased. Many efforts have attempted to reduce inference cost through model compression techniques such as pruning and distillation. However, these techniques either require labeled data, or are time-consuming as they require the compressed model to be retrained to regain accuracy. In this paper, we propose a gradient-free structured pruning framework that uses only unlabeled data. An evaluation on the GLUE and SQuAD benchmarks using BERT_{BASE} and DistilBERT illustrates the effectiveness of the proposed approach. By only using the weights of the pre-trained model and unlabeled data, in a matter of a few minutes on a single GPU, up to 40% of the original FLOP count can be reduced with less than a 4% accuracy loss across all tasks considered.

Modern automatic speech recognition (ASR) models, such as OpenAI's Whisper, rely on deep encoder-decoder architectures, and their encoders are a critical bottleneck for efficient deployment due to high computational intensity. We introduce LiteASR, a low-rank compression scheme for ASR encoders that significantly reduces inference costs while maintaining transcription accuracy. Our approach leverages the strong low-rank properties observed in intermediate activations: by applying principal component analysis (PCA) with a small calibration dataset, we approximate linear transformations with a chain of low-rank matrix multiplications, and further optimize self-attention to work in the reduced dimension. Evaluation results show that our method can compress Whisper large-v3's encoder size by over 50%, matching Whisper medium's size with better transcription accuracy, thereby establishing a new Pareto-optimal frontier of efficiency and performance. The code of LiteASR is available at https://github.com/efeslab/LiteASR.

Recently, long-thought reasoning models achieve strong performance on complex reasoning tasks, but often incur substantial inference overhead, making efficiency a critical concern. Our empirical analysis reveals that the benefit of using Long-CoT varies across problems: while some problems require elaborate reasoning, others show no improvement, or even degraded accuracy. This motivates adaptive reasoning strategies that tailor reasoning depth to the input. However, prior work primarily reduces redundancy within long reasoning paths, limiting exploration of more efficient strategies beyond the Long-CoT paradigm. To address this, we propose a novel two-stage framework for adaptive and efficient reasoning. First, we construct a hybrid reasoning model by merging long and short CoT models to enable diverse reasoning styles. Second, we apply bi-level preference training to guide the model to select suitable reasoning styles (group-level), and prefer concise and correct reasoning within each style group (instance-level). Experiments demonstrate that our method significantly reduces inference costs compared to other baseline approaches, while maintaining performance. Notably, on five mathematical datasets, the average length of reasoning is reduced by more than 50%, highlighting the potential of adaptive strategies to optimize reasoning efficiency in large language models. Our code is coming soon at https://github.com/StarDewXXX/AdaR1

Existing chain-of-thought (CoT) distillation methods can effectively transfer reasoning abilities to base models but suffer from two major limitations: excessive verbosity of reasoning traces and inadequate adaptability to problem difficulty. Long reasoning traces significantly increase inference costs, and uniform-length solutions prevent base models from learning adaptive reasoning strategies. To address these issues, we propose a difficulty-aware prompting (DAP) method to dynamically shorten reasoning traces without performance loss. In our approach, a large teacher model first judges each problem's difficulty and then rewrites its reasoning traces to an appropriate shorter length, yielding concise yet complete reasoning traces. Leveraging the DAP pipeline, we curate a distilled dataset called LiteCoT consisting of 100K concise reasoning examples, with solutions averaging only 720 tokens (an order of magnitude shorter than typical CoTs). Using LiteCoT, we distilled a new family of reasoning models called Liter (1.5B, 7B, and 32B) based on the Qwen2.5 architecture. Experiments show that a student model fine-tuned on just 100K of these difficulty-pruned CoT samples outperforms a model distilled on 800K original Long CoT samples, while significantly reducing training and inference costs. Our method also generalizes well: across 11 diverse benchmarks, the shorter difficulty-aware CoTs achieve equal or better accuracy than Long chains, using far fewer tokens. For example, on the challenging AIME24 exam, our approach reaches 74.2% Pass@1 using only about 5K inference tokens, surpassing other methods that consume many more tokens. Our code and data are available at https://github.com/Evanwu1125/LiteCoT.

Mixture-of-Experts (MoE) models have become a key approach for scaling large language models efficiently by activating only a subset of experts during training and inference. Typically, the number of activated experts presents a trade-off: fewer experts reduce computational costs, while more experts improve performance. Recent studies reveal that not all activated experts contribute equally to model performance, with some providing minimal utility, particularly when finetuning pretrained MoE models for specialized downstream tasks. The co-existence of significant and redundant parameters in experts provides us an opportunity to reduce the number of activated experts while maintaining model performance. In this work, we propose the concept of compressed experts, lightweight modules that serve as compact representations of full experts. Our approach preserves the most important experts while replacing other auxiliary activated experts with compressed experts. The reduction of active parameters significantly lowers inference costs while achieving comparable performance. Extensive experiments on models including Phi-MoE and OLMoE demonstrate that compressed experts recover over 90% of full expert performance across various tasks while reducing more than 30% active parameters and saving 20% in inference costs. This approach enables efficient deployment of MoE models in resource-constrained settings and facilitates scaling to larger models with manageable overhead. Our code is available at https://github.com/yifei-he/Compressed-Experts.

Vision transformers (ViT) usually extract features via forwarding all the tokens in the self-attention layers from top to toe. In this paper, we introduce dynamic token-pass vision transformers (DoViT) for semantic segmentation, which can adaptively reduce the inference cost for images with different complexity. DoViT gradually stops partial easy tokens from self-attention calculation and keeps the hard tokens forwarding until meeting the stopping criteria. We employ lightweight auxiliary heads to make the token-pass decision and divide the tokens into keeping/stopping parts. With a token separate calculation, the self-attention layers are speeded up with sparse tokens and still work friendly with hardware. A token reconstruction module is built to collect and reset the grouped tokens to their original position in the sequence, which is necessary to predict correct semantic masks. We conduct extensive experiments on two common semantic segmentation tasks, and demonstrate that our method greatly reduces about 40% sim 60% FLOPs and the drop of mIoU is within 0.8% for various segmentation transformers. The throughput and inference speed of ViT-L/B are increased to more than 2times on Cityscapes.

With the growth of model and data sizes, a broad effort has been made to design pruning techniques that reduce the resource demand of deep learning pipelines, while retaining model performance. In order to reduce both inference and training costs, a prominent line of work uses low-rank matrix factorizations to represent the network weights. Although able to retain accuracy, we observe that low-rank methods tend to compromise model robustness against adversarial perturbations. By modeling robustness in terms of the condition number of the neural network, we argue that this loss of robustness is due to the exploding singular values of the low-rank weight matrices. Thus, we introduce a robust low-rank training algorithm that maintains the network's weights on the low-rank matrix manifold while simultaneously enforcing approximate orthonormal constraints. The resulting model reduces both training and inference costs while ensuring well-conditioning and thus better adversarial robustness, without compromising model accuracy. This is shown by extensive numerical evidence and by our main approximation theorem that shows the computed robust low-rank network well-approximates the ideal full model, provided a highly performing low-rank sub-network exists.

Large Language Models (LLMs) have made significant strides in problem-solving by incorporating reasoning processes. However, this enhanced reasoning capability results in an increased number of output tokens during inference, leading to higher computational costs. To address this challenge, we propose TwT (Thinking without Tokens), a method that reduces inference-time costs through habitual reasoning distillation with multi-teachers' guidance, while maintaining high performance. Our approach introduces a Habitual Reasoning Distillation method, which internalizes explicit reasoning into the model's habitual behavior through a Teacher-Guided compression strategy inspired by human cognition. Additionally, we propose Dual-Criteria Rejection Sampling (DCRS), a technique that generates a high-quality and diverse distillation dataset using multiple teacher models, making our method suitable for unsupervised scenarios. Experimental results demonstrate that TwT effectively reduces inference costs while preserving superior performance, achieving up to a 13.6% improvement in accuracy with fewer output tokens compared to other distillation methods, offering a highly practical solution for efficient LLM deployment.

Automated log analysis is crucial to ensure high availability and reliability of complex systems. The advent of LLMs in NLP has ushered in a new era of language model-driven automated log analysis, garnering significant interest. Within this field, two primary paradigms based on language models for log analysis have become prominent. Small Language Models (SLMs) follow the pre-train and fine-tune paradigm, focusing on the specific log analysis task through fine-tuning on supervised datasets. On the other hand, LLMs following the in-context learning paradigm, analyze logs by providing a few examples in prompt contexts without updating parameters. Despite their respective strengths, we notice that SLMs are more cost-effective but less powerful, whereas LLMs with large parameters are highly powerful but expensive and inefficient. To trade-off between the performance and inference costs of both models in automated log analysis, this paper introduces an adaptive log analysis framework known as AdaptiveLog, which effectively reduces the costs associated with LLM while ensuring superior results. This framework collaborates an LLM and a small language model, strategically allocating the LLM to tackle complex logs while delegating simpler logs to the SLM. Specifically, to efficiently query the LLM, we propose an adaptive selection strategy based on the uncertainty estimation of the SLM, where the LLM is invoked only when the SLM is uncertain. In addition, to enhance the reasoning ability of the LLM in log analysis tasks, we propose a novel prompt strategy by retrieving similar error-prone cases as the reference, enabling the model to leverage past error experiences and learn solutions from these cases. Extensive experiments demonstrate that AdaptiveLog achieves state-of-the-art results across different tasks, elevating the overall accuracy of log analysis while maintaining cost efficiency.

We introduce tulip agent, an architecture for autonomous LLM-based agents with Create, Read, Update, and Delete access to a tool library containing a potentially large number of tools. In contrast to state-of-the-art implementations, tulip agent does not encode the descriptions of all available tools in the system prompt, which counts against the model's context window, or embed the entire prompt for retrieving suitable tools. Instead, the tulip agent can recursively search for suitable tools in its extensible tool library, implemented exemplarily as a vector store. The tulip agent architecture significantly reduces inference costs, allows using even large tool libraries, and enables the agent to adapt and extend its set of tools. We evaluate the architecture with several ablation studies in a mathematics context and demonstrate its generalizability with an application to robotics. A reference implementation and the benchmark are available at github.com/HRI-EU/tulip_agent.

In the realm of deep learning-based recommendation systems, the increasing computational demands, driven by the growing number of users and items, pose a significant challenge to practical deployment. This challenge is primarily twofold: reducing the model size while effectively learning user and item representations for efficient recommendations. Despite considerable advancements in model compression and architecture search, prevalent approaches face notable constraints. These include substantial additional computational costs from pre-training/re-training in model compression and an extensive search space in architecture design. Additionally, managing complexity and adhering to memory constraints is problematic, especially in scenarios with strict time or space limitations. Addressing these issues, this paper introduces a novel learning paradigm, Dynamic Sparse Learning (DSL), tailored for recommendation models. DSL innovatively trains a lightweight sparse model from scratch, periodically evaluating and dynamically adjusting each weight's significance and the model's sparsity distribution during the training. This approach ensures a consistent and minimal parameter budget throughout the full learning lifecycle, paving the way for "end-to-end" efficiency from training to inference. Our extensive experimental results underline DSL's effectiveness, significantly reducing training and inference costs while delivering comparable recommendation performance.

Recent advancements in reasoning language models have demonstrated remarkable performance in complex tasks, but their extended chain-of-thought reasoning process increases inference overhead. While quantization has been widely adopted to reduce the inference cost of large language models, its impact on reasoning models remains understudied. In this study, we conduct the first systematic study on quantized reasoning models, evaluating the open-sourced DeepSeek-R1-Distilled Qwen and LLaMA families ranging from 1.5B to 70B parameters, and QwQ-32B. Our investigation covers weight, KV cache, and activation quantization using state-of-the-art algorithms at varying bit-widths, with extensive evaluation across mathematical (AIME, MATH-500), scientific (GPQA), and programming (LiveCodeBench) reasoning benchmarks. Our findings reveal that while lossless quantization can be achieved with W8A8 or W4A16 quantization, lower bit-widths introduce significant accuracy risks. We further identify model size, model origin, and task difficulty as critical determinants of performance. Contrary to expectations, quantized models do not exhibit increased output lengths. In addition, strategically scaling the model sizes or reasoning steps can effectively enhance the performance. All quantized models and codes will be open-sourced in https://github.com/ruikangliu/Quantized-Reasoning-Models.

Video generation models hold substantial potential in areas such as filmmaking. However, current video diffusion models need high computational costs and produce suboptimal results due to high complexity of video generation task. In this paper, we propose ConFiner, an efficient high-quality video generation framework that decouples video generation into easier subtasks: structure control and spatial-temporal refinement. It can generate high-quality videos with chain of off-the-shelf diffusion model experts, each expert responsible for a decoupled subtask. During the refinement, we introduce coordinated denoising, which can merge multiple diffusion experts' capabilities into a single sampling. Furthermore, we design ConFiner-Long framework, which can generate long coherent video with three constraint strategies on ConFiner. Experimental results indicate that with only 10\% of the inference cost, our ConFiner surpasses representative models like Lavie and Modelscope across all objective and subjective metrics. And ConFiner-Long can generate high-quality and coherent videos with up to 600 frames.

Sparse Mixture-of-Experts models (MoEs) have recently gained popularity due to their ability to decouple model size from inference efficiency by only activating a small subset of the model parameters for any given input token. As such, sparse MoEs have enabled unprecedented scalability, resulting in tremendous successes across domains such as natural language processing and computer vision. In this work, we instead explore the use of sparse MoEs to scale-down Vision Transformers (ViTs) to make them more attractive for resource-constrained vision applications. To this end, we propose a simplified and mobile-friendly MoE design where entire images rather than individual patches are routed to the experts. We also propose a stable MoE training procedure that uses super-class information to guide the router. We empirically show that our sparse Mobile Vision MoEs (V-MoEs) can achieve a better trade-off between performance and efficiency than the corresponding dense ViTs. For example, for the ViT-Tiny model, our Mobile V-MoE outperforms its dense counterpart by 3.39% on ImageNet-1k. For an even smaller ViT variant with only 54M FLOPs inference cost, our MoE achieves an improvement of 4.66%.

Generative models are transforming creative domains such as music generation, with inference-time strategies like Classifier-Free Guidance (CFG) playing a crucial role. However, CFG doubles inference cost while limiting originality and diversity across generated contents. In this paper, we introduce diversity-rewarded CFG distillation, a novel finetuning procedure that distills the strengths of CFG while addressing its limitations. Our approach optimises two training objectives: (1) a distillation objective, encouraging the model alone (without CFG) to imitate the CFG-augmented predictions, and (2) an RL objective with a diversity reward, promoting the generation of diverse outputs for a given prompt. By finetuning, we learn model weights with the ability to generate high-quality and diverse outputs, without any inference overhead. This also unlocks the potential of weight-based model merging strategies: by interpolating between the weights of two models (the first focusing on quality, the second on diversity), we can control the quality-diversity trade-off at deployment time, and even further boost performance. We conduct extensive experiments on the MusicLM (Agostinelli et al., 2023) text-to-music generative model, where our approach surpasses CFG in terms of quality-diversity Pareto optimality. According to human evaluators, our finetuned-then-merged model generates samples with higher quality-diversity than the base model augmented with CFG. Explore our generations at https://google-research.github.io/seanet/musiclm/diverse_music/.

The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition. Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more. As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs. All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE.

We break the linear link between the layer size and its inference cost by introducing the fast feedforward (FFF) architecture, a log-time alternative to feedforward networks. We demonstrate that FFFs are up to 220x faster than feedforward networks, up to 6x faster than mixture-of-experts networks, and exhibit better training properties than mixtures of experts thanks to noiseless conditional execution. Pushing FFFs to the limit, we show that they can use as little as 1% of layer neurons for inference in vision transformers while preserving 94.2% of predictive performance.

We show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve a 75% relative improvement in a challenging domain transfer scenario.

Deploying transformer models in practice is challenging due to their inference cost, which scales quadratically with input sequence length. To address this, we present a novel Learned Token Pruning (LTP) method which adaptively removes unimportant tokens as an input sequence passes through transformer layers. In particular, LTP prunes tokens with an attention score below a threshold value which is learned for each layer during training. Our threshold-based method allows the length of the pruned sequence to vary adaptively based on the input sequence, and avoids algorithmically expensive operations such as top-k token selection. We extensively test the performance of LTP on GLUE tasks and show that our method outperforms the prior state-of-the-art token pruning methods by up to ~2.5% higher accuracy with the same amount of FLOPs. In particular, LTP achieves up to 2.1x FLOPs reduction with less than 1% accuracy drop, which results in up to 1.9x and 2.0x throughput improvement on Intel Haswell CPUs and NVIDIA V100 GPUs, respectively. Furthermore, we demonstrate that LTP is more robust than prior methods to variations on input sentence lengths. Our code has been developed in PyTorch and has been open-sourced.

Diffusion Models (DM) have democratized AI image generation through an iterative denoising process. Quantization is a major technique to alleviate the inference cost and reduce the size of DM denoiser networks. However, as denoisers evolve from variants of convolutional U-Nets toward newer Transformer architectures, it is of growing importance to understand the quantization sensitivity of different weight layers, operations and architecture types to performance. In this work, we address this challenge with Qua^2SeDiMo, a mixed-precision Post-Training Quantization framework that generates explainable insights on the cost-effectiveness of various model weight quantization methods for different denoiser operation types and block structures. We leverage these insights to make high-quality mixed-precision quantization decisions for a myriad of diffusion models ranging from foundational U-Nets to state-of-the-art Transformers. As a result, Qua^2SeDiMo can construct 3.4-bit, 3.9-bit, 3.65-bit and 3.7-bit weight quantization on PixArt-{alpha}, PixArt-{Sigma}, Hunyuan-DiT and SDXL, respectively. We further pair our weight-quantization configurations with 6-bit activation quantization and outperform existing approaches in terms of quantitative metrics and generative image quality.

Leading open-source large language models (LLMs) such as Llama-3.1-Instruct-405B are extremely capable at generating text, answering questions, and solving a variety of natural language understanding tasks. However, they incur higher inference cost and latency compared to smaller LLMs. Knowledge distillation provides a way to use outputs from these large, capable teacher models to train smaller student models which can be used for inference at lower cost and latency, while retaining comparable accuracy. We investigate the efficacy of distillation using the Llama-3.1-405B-Instruct teacher and the smaller Llama-3.1-8B-Instruct and Llama-3.1-70B-Instruct student models. Contributions of this work include (a) We evaluate the generalizability of distillation with the above Llama-3.1 teacher-student pairs across different tasks and datasets (b) We show that using synthetic data during distillation significantly improves the accuracy of 8B and 70B models, and when used with reasoning chains, even matches or surpasses the zero-shot accuracy of 405B model on some datasets (c) We empirically show that distillation enables 8B and 70B models to internalize 405B's reasoning ability by using only standard fine-tuning (without customizing any loss function). This allows cost and latency-efficient student model inference. (d) We show pitfalls in evaluation of distillation, and present task-specific evaluation, including both human and LLM-grading, and ground-truth based traditional accuracy benchmarks. This methodical study brings out the fundamental importance of synthetic data quality in knowledge distillation, and of combining multiple, task-specific ways of accuracy and quality evaluation in assessing the effectiveness of distillation.

Recent advancements in timestep-distilled diffusion models have enabled high-quality image generation that rivals non-distilled multi-step models, but with significantly fewer inference steps. While such models are attractive for applications due to the low inference cost and latency, fine-tuning them with a naive diffusion objective would result in degraded and blurry outputs. An intuitive alternative is to repeat the diffusion distillation process with a fine-tuned teacher model, which produces good results but is cumbersome and computationally intensive; the distillation training usually requires magnitude higher of training compute compared to fine-tuning for specific image styles. In this paper, we present an algorithm named pairwise sample optimization (PSO), which enables the direct fine-tuning of an arbitrary timestep-distilled diffusion model. PSO introduces additional reference images sampled from the current time-step distilled model, and increases the relative likelihood margin between the training images and reference images. This enables the model to retain its few-step generation ability, while allowing for fine-tuning of its output distribution. We also demonstrate that PSO is a generalized formulation which can be flexibly extended to both offline-sampled and online-sampled pairwise data, covering various popular objectives for diffusion model preference optimization. We evaluate PSO in both preference optimization and other fine-tuning tasks, including style transfer and concept customization. We show that PSO can directly adapt distilled models to human-preferred generation with both offline and online-generated pairwise preference image data. PSO also demonstrates effectiveness in style transfer and concept customization by directly tuning timestep-distilled diffusion models.

Self-ensembling techniques with diverse reasoning paths such as Self-Consistency have demonstrated remarkable performance gains in text generation with Large Language Models (LLMs). However, such techniques depend on the availability of an accurate answer extraction process to aggregate across multiple outputs. Moreover, they acquire higher inference cost, in comparison to Greedy Decoding, due to generation of relatively higher number of output tokens. Research has shown that the free form text outputs from Self-Consistency can be aggregated reliably using LLMs to produce the final output. Additionally, recent advancements in LLM inference have demonstrated that usage of diverse exemplars in prompts have the ability to induce diversity in the LLM outputs. Such proven techniques can be easily extended to self-ensembling based approaches to achieve enhanced results in text generation. In this paper, we introduce PEDAL (Prompts based on Exemplar Diversity Aggregated using LLMs), a hybrid self-ensembling approach, that combines the strengths of diverse exemplar based prompts and LLM based aggregation to achieve improvement in overall performance. On the publicly available SVAMP and ARC datasets, our experiments reveal that PEDAL can achieve better accuracy than Greedy Decoding based strategies with lower inference cost compared to Self Consistency based approaches.

The recent emergence of Large Language Models based on the Transformer architecture has enabled dramatic advancements in the field of Natural Language Processing. However, these models have long inference latency, which limits their deployment and makes them prohibitively expensive for various real-time applications. The inference latency is further exacerbated by autoregressive generative tasks, as models need to run iteratively to generate tokens sequentially without leveraging token-level parallelization. To address this, we propose Big Little Decoder (BiLD), a framework that can improve inference efficiency and latency for a wide range of text generation applications. The BiLD framework contains two models with different sizes that collaboratively generate text. The small model runs autoregressively to generate text with a low inference cost, and the large model is only invoked occasionally to refine the small model's inaccurate predictions in a non-autoregressive manner. To coordinate the small and large models, BiLD introduces two simple yet effective policies: (1) the fallback policy that determines when to hand control over to the large model; and (2) the rollback policy that determines when the large model needs to correct the small model's inaccurate predictions. To evaluate our framework across different tasks and models, we apply BiLD to various text generation scenarios encompassing machine translation on IWSLT 2017 De-En and WMT 2014 De-En, and summarization on XSUM and CNN/DailyMail. On an NVIDIA T4 GPU, our framework achieves a speedup of up to 2.12x speedup with minimal generation quality degradation. Furthermore, our framework is fully plug-and-play and can be applied without any modifications in the training process or model architecture. Our code is open-sourced

Network pruning is widely used for reducing the heavy inference cost of deep models in low-resource settings. A typical pruning algorithm is a three-stage pipeline, i.e., training (a large model), pruning and fine-tuning. During pruning, according to a certain criterion, redundant weights are pruned and important weights are kept to best preserve the accuracy. In this work, we make several surprising observations which contradict common beliefs. For all state-of-the-art structured pruning algorithms we examined, fine-tuning a pruned model only gives comparable or worse performance than training that model with randomly initialized weights. For pruning algorithms which assume a predefined target network architecture, one can get rid of the full pipeline and directly train the target network from scratch. Our observations are consistent for multiple network architectures, datasets, and tasks, which imply that: 1) training a large, over-parameterized model is often not necessary to obtain an efficient final model, 2) learned "important" weights of the large model are typically not useful for the small pruned model, 3) the pruned architecture itself, rather than a set of inherited "important" weights, is more crucial to the efficiency in the final model, which suggests that in some cases pruning can be useful as an architecture search paradigm. Our results suggest the need for more careful baseline evaluations in future research on structured pruning methods. We also compare with the "Lottery Ticket Hypothesis" (Frankle & Carbin 2019), and find that with optimal learning rate, the "winning ticket" initialization as used in Frankle & Carbin (2019) does not bring improvement over random initialization.

Recently, large reasoning models have achieved impressive performance on various tasks by employing human-like deep thinking. However, the lengthy thinking process substantially increases inference overhead, making efficiency a critical bottleneck. In this work, we first demonstrate that NoThinking, which prompts the reasoning model to skip thinking and directly generate the final solution, is a better choice for relatively simple tasks in terms of both performance and efficiency. Motivated by this, we propose AdaptThink, a novel RL algorithm to teach reasoning models to choose the optimal thinking mode adaptively based on problem difficulty. Specifically, AdaptThink features two core components: (1) a constrained optimization objective that encourages the model to choose NoThinking while maintaining the overall performance; (2) an importance sampling strategy that balances Thinking and NoThinking samples during on-policy training, thereby enabling cold start and allowing the model to explore and exploit both thinking modes throughout the training process. Our experiments indicate that AdaptThink significantly reduces the inference costs while further enhancing performance. Notably, on three math datasets, AdaptThink reduces the average response length of DeepSeek-R1-Distill-Qwen-1.5B by 53% and improves its accuracy by 2.4%, highlighting the promise of adaptive thinking-mode selection for optimizing the balance between reasoning quality and efficiency. Our codes and models are available at https://github.com/THU-KEG/AdaptThink.

Quantizing model weights is critical for reducing the communication and inference costs of large models. However, quantizing models -- especially to low precisions like int4 or int2 -- requires a trade-off in model quality; int2, in particular, is known to severely degrade model quality. Consequently, practitioners are often forced to maintain multiple models with different quantization levels or serve a single model that best satisfies the quality-latency trade-off. On the other hand, integer data types, such as int8, inherently possess a nested (Matryoshka) structure where smaller bit-width integers, like int4 or int2, are nested within the most significant bits. This paper proposes Matryoshka Quantization (MatQuant), a novel multi-scale quantization technique that addresses the challenge of needing multiple quantized models. It allows training and maintaining just one model, which can then be served at different precision levels. Furthermore, due to the co-training and co-distillation regularization provided by MatQuant, the int2 precision models extracted by MatQuant can be up to 10% more accurate than standard int2 quantization (using techniques like QAT or OmniQuant). This represents significant progress in model quantization, demonstrated by the fact that, with the same recipe, an int2 FFN-quantized Gemma-2 9B model is more accurate than an int8 FFN-quantized Gemma-2 2B model.

This paper introduces MultiBooth, a novel and efficient technique for multi-concept customization in image generation from text. Despite the significant advancements in customized generation methods, particularly with the success of diffusion models, existing methods often struggle with multi-concept scenarios due to low concept fidelity and high inference cost. MultiBooth addresses these issues by dividing the multi-concept generation process into two phases: a single-concept learning phase and a multi-concept integration phase. During the single-concept learning phase, we employ a multi-modal image encoder and an efficient concept encoding technique to learn a concise and discriminative representation for each concept. In the multi-concept integration phase, we use bounding boxes to define the generation area for each concept within the cross-attention map. This method enables the creation of individual concepts within their specified regions, thereby facilitating the formation of multi-concept images. This strategy not only improves concept fidelity but also reduces additional inference cost. MultiBooth surpasses various baselines in both qualitative and quantitative evaluations, showcasing its superior performance and computational efficiency. Project Page: https://multibooth.github.io/

In deep learning, mixture-of-experts (MoE) activates one or few experts (sub-networks) on a per-sample or per-token basis, resulting in significant computation reduction. The recently proposed patch-level routing in MoE (pMoE) divides each input into n patches (or tokens) and sends l patches (lll n) to each expert through prioritized routing. pMoE has demonstrated great empirical success in reducing training and inference costs while maintaining test accuracy. However, the theoretical explanation of pMoE and the general MoE remains elusive. Focusing on a supervised classification task using a mixture of two-layer convolutional neural networks (CNNs), we show for the first time that pMoE provably reduces the required number of training samples to achieve desirable generalization (referred to as the sample complexity) by a factor in the polynomial order of n/l, and outperforms its single-expert counterpart of the same or even larger capacity. The advantage results from the discriminative routing property, which is justified in both theory and practice that pMoE routers can filter label-irrelevant patches and route similar class-discriminative patches to the same expert. Our experimental results on MNIST, CIFAR-10, and CelebA support our theoretical findings on pMoE's generalization and show that pMoE can avoid learning spurious correlations.

Backward compatibility of model predictions is a desired property when updating a machine learning driven application. It allows to seamlessly improve the underlying model without introducing regression bugs. In classification tasks these bugs occur in the form of negative flips. This means an instance that was correctly classified by the old model is now classified incorrectly by the updated model. This has direct negative impact on the user experience of such systems e.g. a frequently used voice assistant query is suddenly misclassified. A common reason to update the model is when new training data becomes available and needs to be incorporated. Simply retraining the model with the updated data introduces the unwanted negative flips. We study the problem of regression during data updates and propose Backward Compatible Weight Interpolation (BCWI). This method interpolates between the weights of the old and new model and we show in extensive experiments that it reduces negative flips without sacrificing the improved accuracy of the new model. BCWI is straight forward to implement and does not increase inference cost. We also explore the use of importance weighting during interpolation and averaging the weights of multiple new models in order to further reduce negative flips.

Large language models (LLMs), both proprietary and open-source, have demonstrated remarkable capabilities across various natural language processing tasks. However, they face significant limitations in legal reasoning tasks. Proprietary models introduce data privacy risks and high inference costs, while open-source models underperform due to insufficient legal domain training data. To address these limitations, we study data generation for legal reasoning to improve the legal reasoning performance of open-source LLMs with the help of proprietary LLMs. This is challenging due to the lack of legal knowledge in proprietary LLMs and the difficulty in verifying the generated data. We propose KgDG, a knowledge-guided data generation framework for legal reasoning. Our framework enables leveraging legal knowledge to enhance generation diversity and introduces a refinement and verification process to ensure the quality of generated data. Moreover, we expand the generated dataset to further enhance the LLM reasoning capabilities. Using KgDG, we create a synthetic legal reasoning dataset containing 50K high-quality examples. Our trained model LawGPT outperforms existing legal-specific LLMs and achieves performance comparable to proprietary LLMs, demonstrating the effectiveness of KgDG and LawGPT. Our code and resources is publicly available at https://anonymous.4open.science/r/KgDG-45F5 .

The o1-Like LLMs are transforming AI by simulating human cognitive processes, but their performance in multilingual machine translation (MMT) remains underexplored. This study examines: (1) how o1-Like LLMs perform in MMT tasks and (2) what factors influence their translation quality. We evaluate multiple o1-Like LLMs and compare them with traditional models like ChatGPT and GPT-4o. Results show that o1-Like LLMs establish new multilingual translation benchmarks, with DeepSeek-R1 surpassing GPT-4o in contextless tasks. They demonstrate strengths in historical and cultural translation but exhibit a tendency for rambling issues in Chinese-centric outputs. Further analysis reveals three key insights: (1) High inference costs and slower processing speeds make complex translation tasks more resource-intensive. (2) Translation quality improves with model size, enhancing commonsense reasoning and cultural translation. (3) The temperature parameter significantly impacts output quality-lower temperatures yield more stable and accurate translations, while higher temperatures reduce coherence and precision.

Recent advances show that large language models (LLMs) generalize strong performance across different natural language benchmarks. However, the large size of LLMs makes training and inference expensive and impractical to run in resource-limited settings. This paper introduces a new approach called fine-tuning stacks of language models (FSLM), which involves stacking small language models (SLM) as an alternative to LLMs. By fine-tuning each SLM to perform a specific task, this approach breaks down high level reasoning into multiple lower-level steps that specific SLMs are responsible for. As a result, FSLM allows for lower training and inference costs, and also improves model interpretability as each SLM communicates with the subsequent one through natural language. By evaluating FSLM on common natural language benchmarks, this paper highlights promising early results toward generalizable performance using FSLM as a cost-effective alternative to LLMs.

Recent advancements in large language models (LLMs) have raised concerns about inference costs, increasing the need for research into model compression. While knowledge distillation (KD) is a prominent method for this, research on KD for generative language models like LLMs is relatively sparse, and the approach of distilling student-friendly knowledge, which has shown promising performance in KD for classification models, remains unexplored in generative language models. To explore this approach, we propose PromptKD, a simple yet effective method that utilizes prompt tuning - for the first time in KD - to enable generative language models to transfer student-friendly knowledge. Unlike previous works in classification that require fine-tuning the entire teacher model for extracting student-friendly knowledge, PromptKD achieves similar effects by adding a small number of prompt tokens and tuning only the prompt with student guidance. Extensive experiments on instruction-following datasets using the GPT-2 model family show that PromptKD achieves state-of-the-art performance while adding only 0.0007% of the teacher's parameters as prompts. Further analysis suggests that distilling student-friendly knowledge alleviates exposure bias effectively throughout the entire training process, leading to performance enhancements.

Deep ensemble is a simple yet powerful way to improve the performance of deep neural networks. Under this motivation, recent works on mode connectivity have shown that parameters of ensembles are connected by low-loss subspaces, and one can efficiently collect ensemble parameters in those subspaces. While this provides a way to efficiently train ensembles, for inference, multiple forward passes should still be executed using all the ensemble parameters, which often becomes a serious bottleneck for real-world deployment. In this work, we propose a novel framework to reduce such costs. Given a low-loss subspace connecting two modes of a neural network, we build an additional neural network that predicts the output of the original neural network evaluated at a certain point in the low-loss subspace. The additional neural network, which we call a "bridge", is a lightweight network that takes minimal features from the original network and predicts outputs for the low-loss subspace without forward passes through the original network. We empirically demonstrate that we can indeed train such bridge networks and significantly reduce inference costs with the help of bridge networks.

The computing cost and memory demand of deep learning pipelines have grown fast in recent years and thus a variety of pruning techniques have been developed to reduce model parameters. The majority of these techniques focus on reducing inference costs by pruning the network after a pass of full training. A smaller number of methods address the reduction of training costs, mostly based on compressing the network via low-rank layer factorizations. Despite their efficiency for linear layers, these methods fail to effectively handle convolutional filters. In this work, we propose a low-parametric training method that factorizes the convolutions into tensor Tucker format and adaptively prunes the Tucker ranks of the convolutional kernel during training. Leveraging fundamental results from geometric integration theory of differential equations on tensor manifolds, we obtain a robust training algorithm that provably approximates the full baseline performance and guarantees loss descent. A variety of experiments against the full model and alternative low-rank baselines are implemented, showing that the proposed method drastically reduces the training costs, while achieving high performance, comparable to or better than the full baseline, and consistently outperforms competing low-rank approaches.

We present Synergy Aware Forgetting Ensemble (SAFE), a method to adapt large models on a diverse collection of data while minimizing the expected cost to remove the influence of training samples from the trained model. This process, also known as selective forgetting or unlearning, is often conducted by partitioning a dataset into shards, training fully independent models on each, then ensembling the resulting models. Increasing the number of shards reduces the expected cost to forget but at the same time it increases inference cost and reduces the final accuracy of the model since synergistic information between samples is lost during the independent model training. Rather than treating each shard as independent, SAFE introduces the notion of a shard graph, which allows incorporating limited information from other shards during training, trading off a modest increase in expected forgetting cost with a significant increase in accuracy, all while still attaining complete removal of residual influence after forgetting. SAFE uses a lightweight system of adapters which can be trained while reusing most of the computations. This allows SAFE to be trained on shards an order-of-magnitude smaller than current state-of-the-art methods (thus reducing the forgetting costs) while also maintaining high accuracy, as we demonstrate empirically on fine-grained computer vision datasets.

Existing methods enhance the training of detection transformers by incorporating an auxiliary one-to-many assignment. In this work, we treat the model as a multi-task framework, simultaneously performing one-to-one and one-to-many predictions. We investigate the roles of each component in the transformer decoder across these two training targets, including self-attention, cross-attention, and feed-forward network. Our empirical results demonstrate that any independent component in the decoder can effectively learn both targets simultaneously, even when other components are shared. This finding leads us to propose a multi-route training mechanism, featuring a primary route for one-to-one prediction and two auxiliary training routes for one-to-many prediction. We enhance the training mechanism with a novel instructive self-attention that dynamically and flexibly guides object queries for one-to-many prediction. The auxiliary routes are removed during inference, ensuring no impact on model architecture or inference cost. We conduct extensive experiments on various baselines, achieving consistent improvements as shown in Figure 1. Project page: https://visual-ai.github.io/mrdetr

Transformer models struggle with long-context inference due to their quadratic time and linear memory complexity. Recurrent Memory Transformers (RMTs) offer a solution by reducing the asymptotic cost to linear time and constant memory usage. However, their memory update mechanism leads to sequential execution, causing a performance bottleneck. We introduce Diagonal Batching, a scheduling scheme that unlocks parallelism across segments in RMTs while preserving exact recurrence. This approach eliminates the sequential constraint, enabling efficient GPU inference even for single long-context inputs without complex batching and pipelining techniques. Because the technique is purely a run-time computation reordering, existing RMT models adopt it with no retraining. Applied to a LLaMA-1B ARMT model, Diagonal Batching yields a 3.3x speedup over standard full-attention LLaMA-1B and a 1.8x speedup over the sequential RMT implementation on 131,072-token sequences. By removing sequential bottleneck, Diagonal Batching reduces inference cost and latency, thereby strengthening RMTs as a practical solution for real-world, long-context applications.

Mixture-of-Experts (MoE) architectures offer a general solution to the high inference costs of large language models (LLMs) via sparse routing, bringing faster and more accurate models, at the cost of massive parameter counts. For example, the SwitchTransformer-c2048 model has 1.6 trillion parameters, requiring 3.2TB of accelerator memory to run efficiently, which makes practical deployment challenging and expensive. In this paper, we present a solution to this memory problem, in form of a new compression and execution framework called QMoE. Specifically, QMoE consists of a scalable algorithm which accurately compresses trillion-parameter MoEs to less than 1 bit per parameter, in a custom format co-designed with bespoke GPU decoding kernels to facilitate efficient end-to-end compressed inference, with minor runtime overheads relative to uncompressed execution. Concretely, QMoE can compress the 1.6 trillion parameter SwitchTransformer-c2048 model to less than 160GB (20x compression, 0.8 bits per parameter) at only minor accuracy loss, in less than a day on a single GPU. This enables, for the first time, the execution of a trillion-parameter model on affordable commodity hardware, like a single server with 4x NVIDIA A6000 or 8x NVIDIA 3090 GPUs, at less than 5% runtime overhead relative to ideal uncompressed inference. The source code and compressed models are available at github.com/IST-DASLab/qmoe.

Knowledge distillation (KD) is widely used for compressing a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, current KD methods for auto-regressive sequence models suffer from distribution mismatch between output sequences seen during training and those generated by the student during inference. To address this issue, we introduce Generalized Knowledge Distillation (GKD). Instead of solely relying on a fixed set of output sequences, GKD trains the student on its self-generated output sequences by leveraging feedback from the teacher on such sequences. Unlike supervised KD approaches, GKD also offers the flexibility to employ alternative loss functions between the student and teacher, which can be useful when the student lacks the expressivity to mimic the teacher's distribution. Furthermore, GKD facilitates the seamless integration of distillation with RL fine-tuning (RLHF). We demonstrate the efficacy of GKD for distilling auto-regressive language models on summarization, translation, and arithmetic reasoning tasks, and task-agnostic distillation for instruction-tuning.

We introduce the Falcon series: 7B, 40B, and 180B parameters causal decoder-only models trained on a diverse high-quality corpora predominantly assembled from web data. The largest model, Falcon-180B, has been trained on over 3.5 trillion tokens of text--the largest openly documented pretraining run. Falcon-180B significantly outperforms models such as PaLM or Chinchilla, and improves upon concurrently developed models such as LLaMA 2 or Inflection-1. It nears the performance of PaLM-2-Large at a reduced pretraining and inference cost, making it, to our knowledge, one of the three best language models in the world along with GPT-4 and PaLM-2-Large. We report detailed evaluations, as well as a deep dive into the methods and custom tooling employed to pretrain Falcon. Notably, we report on our custom distributed training codebase, allowing us to efficiently pretrain these models on up to 4,096 A100s on cloud AWS infrastructure with limited interconnect. We release a 600B tokens extract of our web dataset, as well as the Falcon-7/40/180B models under a permissive license to foster open-science and accelerate the development of an open ecosystem of large language models.

Large language models (LLMs) are increasingly evaluated on single-answer multiple-choice tasks, yet many real-world problems require identifying all correct answers from a set of options. This capability remains underexplored. We introduce SATA-BENCH, the first dedicated benchmark for evaluating LLMs on Select All That Apply (SATA) questions across diverse domains, including reading comprehension, law, and biomedicine. Our evaluation of 27 open-source and proprietary models reveals a significant gap: even the strongest model achieves only 41.8% exact match, exposing LLMs' inability to reliably identify all correct answers. We find that this weakness stems from two core challenges: selection bias - models favor certain choices regardless of content, and count bias - models fail to predict the correct number of answers. To address these issues, we propose Choice Funnel, a decoding strategy that combines token debiasing with adaptive thresholding to guide models toward complete and accurate selections. Choice Funnel achieves up to 29% higher exact match than competitive baselines while reducing inference cost by over 64%. Our findings expose fundamental limitations in current LLMs and introduce a new framework for diagnosing and improving multi-answer reasoning. We release SATA-BENCH and Choice Funnel to promote LLM development for robust decision-making in realistic, multi-answer applications.

Deployment of Transformer models on edge devices is becoming increasingly challenging due to the exponentially growing inference cost that scales quadratically with the number of tokens in the input sequence. Token pruning is an emerging solution to address this challenge due to its ease of deployment on various Transformer backbones. However, most token pruning methods require computationally expensive fine-tuning, which is undesirable in many edge deployment cases. In this work, we propose Zero-TPrune, the first zero-shot method that considers both the importance and similarity of tokens in performing token pruning. It leverages the attention graph of pre-trained Transformer models to produce an importance distribution for tokens via our proposed Weighted Page Rank (WPR) algorithm. This distribution further guides token partitioning for efficient similarity-based pruning. Due to the elimination of the fine-tuning overhead, Zero-TPrune can prune large models at negligible computational cost, switch between different pruning configurations at no computational cost, and perform hyperparameter tuning efficiently. We evaluate the performance of Zero-TPrune on vision tasks by applying it to various vision Transformer backbones and testing them on ImageNet. Without any fine-tuning, Zero-TPrune reduces the FLOPs cost of DeiT-S by 34.7\% and improves its throughput by 45.3\% with only 0.4\% accuracy loss. Compared with state-of-the-art pruning methods that require fine-tuning, Zero-TPrune not only eliminates the need for fine-tuning after pruning but also does so with only 0.1\% accuracy loss. Compared with state-of-the-art fine-tuning-free pruning methods, Zero-TPrune reduces accuracy loss by up to 49\% with the same or higher throughput.

Task-oriented dialogue systems are broadly used in virtual assistants and other automated services, providing interfaces between users and machines to facilitate specific tasks. Nowadays, task-oriented dialogue systems have greatly benefited from pre-trained language models (PLMs). However, their task-solving performance is constrained by the inherent capacities of PLMs, and scaling these models is expensive and complex as the model size becomes larger. To address these challenges, we propose Soft Mixture-of-Expert Task-Oriented Dialogue system (SMETOD) which leverages an ensemble of Mixture-of-Experts (MoEs) to excel at subproblems and generate specialized outputs for task-oriented dialogues. SMETOD also scales up a task-oriented dialogue system with simplicity and flexibility while maintaining inference efficiency. We extensively evaluate our model on three benchmark functionalities: intent prediction, dialogue state tracking, and dialogue response generation. Experimental results demonstrate that SMETOD achieves state-of-the-art performance on most evaluated metrics. Moreover, comparisons against existing strong baselines show that SMETOD has a great advantage in the cost of inference and correctness in problem-solving.

Large language models (LLMs) have demonstrated impressive abilities in various domains while the inference cost is expensive. The state-of-the-art methods use 2-bit quantization for mainstream LLMs. However, challenges still exist: (1) Nonnegligible accuracy loss for 2-bit quantization. Weights are quantized by groups, while the ranges of weights are large in some groups, resulting in large quantization errors and nonnegligible accuracy loss (e.g. >3% for Llama2-7b with 2-bit quantization in GPTQ and Greenbit). (2) Limited accuracy improvement by adding 4-bit weights. Increasing 10% extra average bit more 4-bit weights only leads to <0.5% accuracy improvement on a quantized Llama2-7b. (3) Time-consuming dequantization operations on GPUs. The dequantization operations lead to >50% execution time, hindering the potential of reducing LLM inference cost. To tackle these challenges, we propose the following techniques: (1) We only quantize a small fraction of groups with the larger range using 4-bit with memory alignment consideration on GPUs.(2) We design the asynchronous dequantization on GPUs, leading to up to 3.92X speedup. We conduct extensive experiments on different model sizes. We achieve 2.85-bit for each weight and the end-to-end speedup for Llama2-7b is 1.74X over the original model, and we reduce both runtime cost and hardware cost by up to 2.70X and 2.81X with less GPU requirements.

To build an artificial neural network like the biological intelligence system, recent works have unified numerous tasks into a generalist model, which can process various tasks with shared parameters and do not have any task-specific modules. While generalist models achieve promising results on various benchmarks, they have performance degradation on some tasks compared with task-specialized models. In this work, we find that interference among different tasks and modalities is the main factor to this phenomenon. To mitigate such interference, we introduce the Conditional Mixture-of-Experts (Conditional MoEs) to generalist models. Routing strategies under different levels of conditions are proposed to take both the training/inference cost and generalization ability into account. By incorporating the proposed Conditional MoEs, the recently proposed generalist model Uni-Perceiver can effectively mitigate the interference across tasks and modalities, and achieves state-of-the-art results on a series of downstream tasks via prompt tuning on 1% of downstream data. Moreover, the introduction of Conditional MoEs still holds the generalization ability of generalist models to conduct zero-shot inference on new tasks, e.g., video-text retrieval and video caption. Code and pre-trained generalist models shall be released.

An emerging recipe for achieving state-of-the-art effectiveness in neural document re-ranking involves utilizing large pre-trained language models - e.g., BERT - to evaluate all individual passages in the document and then aggregating the outputs by pooling or additional Transformer layers. A major drawback of this approach is high query latency due to the cost of evaluating every passage in the document with BERT. To make matters worse, this high inference cost and latency varies based on the length of the document, with longer documents requiring more time and computation. To address this challenge, we adopt an intra-document cascading strategy, which prunes passages of a candidate document using a less expensive model, called ESM, before running a scoring model that is more expensive and effective, called ETM. We found it best to train ESM (short for Efficient Student Model) via knowledge distillation from the ETM (short for Effective Teacher Model) e.g., BERT. This pruning allows us to only run the ETM model on a smaller set of passages whose size does not vary by document length. Our experiments on the MS MARCO and TREC Deep Learning Track benchmarks suggest that the proposed Intra-Document Cascaded Ranking Model (IDCM) leads to over 400% lower query latency by providing essentially the same effectiveness as the state-of-the-art BERT-based document ranking models.

Sparse Mixtures of Experts (SMoE) scales model capacity without significant increases in training and inference costs, but exhibits the following two issues: (1) Low expert activation, where only a small subset of experts are activated for optimization. (2) Lacking fine-grained analytical capabilities for multiple semantic concepts within individual tokens. We propose Multi-Head Mixture-of-Experts (MH-MoE), which employs a multi-head mechanism to split each token into multiple sub-tokens. These sub-tokens are then assigned to and processed by a diverse set of experts in parallel, and seamlessly reintegrated into the original token form. The multi-head mechanism enables the model to collectively attend to information from various representation spaces within different experts, while significantly enhances expert activation, thus deepens context understanding and alleviate overfitting. Moreover, our MH-MoE is straightforward to implement and decouples from other SMoE optimization methods, making it easy to integrate with other SMoE models for enhanced performance. Extensive experimental results across three tasks: English-focused language modeling, Multi-lingual language modeling and Masked multi-modality modeling tasks, demonstrate the effectiveness of MH-MoE.

Low-precision training is considered an effective strategy for reducing both training and downstream inference costs. Previous scaling laws for precision mainly focus on integer quantization, which pay less attention to the constituents in floating-point quantization and thus cannot well fit the LLM losses in this scenario. In contrast, while floating-point quantization training is more commonly implemented in production, the research on it has been relatively superficial. In this paper, we thoroughly explore the effects of floating-point quantization targets, exponent bits, mantissa bits, and the calculation granularity of the scaling factor in floating-point quantization training performance of LLM models. While presenting an accurate floating-point quantization unified scaling law, we also provide valuable suggestions for the community: (1) Exponent bits contribute slightly more to the model performance than mantissa bits. We provide the optimal exponent-mantissa bit ratio for different bit numbers, which is available for future reference by hardware manufacturers; (2) We discover the formation of the critical data size in low-precision LLM training. Too much training data exceeding the critical data size will inversely bring in degradation of LLM performance; (3) The optimal floating-point quantization precision is directly proportional to the computational power, but within a wide computational power range, we estimate that the best cost-performance precision lies between 4-8 bits.

Deep learning for time series forecasting has seen significant advancements over the past decades. However, despite the success of large-scale pre-training in language and vision domains, pre-trained time series models remain limited in scale and operate at a high cost, hindering the development of larger capable forecasting models in real-world applications. In response, we introduce Time-MoE, a scalable and unified architecture designed to pre-train larger, more capable forecasting foundation models while reducing inference costs. By leveraging a sparse mixture-of-experts (MoE) design, Time-MoE enhances computational efficiency by activating only a subset of networks for each prediction, reducing computational load while maintaining high model capacity. This allows Time-MoE to scale effectively without a corresponding increase in inference costs. Time-MoE comprises a family of decoder-only transformer models that operate in an auto-regressive manner and support flexible forecasting horizons with varying input context lengths. We pre-trained these models on our newly introduced large-scale data Time-300B, which spans over 9 domains and encompassing over 300 billion time points. For the first time, we scaled a time series foundation model up to 2.4 billion parameters, achieving significantly improved forecasting precision. Our results validate the applicability of scaling laws for training tokens and model size in the context of time series forecasting. Compared to dense models with the same number of activated parameters or equivalent computation budgets, our models consistently outperform them by large margin. These advancements position Time-MoE as a state-of-the-art solution for tackling real-world time series forecasting challenges with superior capability, efficiency, and flexibility.

This paper presents a new mechanism to facilitate the training of mask transformers for efficient panoptic segmentation, democratizing its deployment. We observe that due to its high complexity, the training objective of panoptic segmentation will inevitably lead to much higher false positive penalization. Such unbalanced loss makes the training process of the end-to-end mask-transformer based architectures difficult, especially for efficient models. In this paper, we present ReMaX that adds relaxation to mask predictions and class predictions during training for panoptic segmentation. We demonstrate that via these simple relaxation techniques during training, our model can be consistently improved by a clear margin without any extra computational cost on inference. By combining our method with efficient backbones like MobileNetV3-Small, our method achieves new state-of-the-art results for efficient panoptic segmentation on COCO, ADE20K and Cityscapes. Code and pre-trained checkpoints will be available at https://github.com/google-research/deeplab2.

Despite their impressive performance on diverse tasks, large language models (LMs) still struggle with tasks requiring rich world knowledge, implying the limitations of relying solely on their parameters to encode a wealth of world knowledge. This paper aims to understand LMs' strengths and limitations in memorizing factual knowledge, by conducting large-scale knowledge probing experiments of 10 models and 4 augmentation methods on PopQA, our new open-domain QA dataset with 14k questions. We find that LMs struggle with less popular factual knowledge, and that scaling fails to appreciably improve memorization of factual knowledge in the long tail. We then show that retrieval-augmented LMs largely outperform orders of magnitude larger LMs, while unassisted LMs remain competitive in questions about high-popularity entities. Based on those findings, we devise a simple, yet effective, method for powerful and efficient retrieval-augmented LMs, which retrieves non-parametric memories only when necessary. Experimental results show that this significantly improves models' performance while reducing the inference costs.

We introduce a novel framework for analyzing sorting algorithms in pairwise ranking prompting (PRP), re-centering the cost model around LLM inferences rather than traditional pairwise comparisons. While classical metrics based on comparison counts have traditionally been used to gauge efficiency, our analysis reveals that expensive LLM inferences overturn these predictions; accordingly, our framework encourages strategies such as batching and caching to mitigate inference costs. We show that algorithms optimal in the classical setting can lose efficiency when LLM inferences dominate the cost under certain optimizations.

This paper presents a simple, effective, and cost-efficient strategy to improve LLM performance by scaling test-time compute. Our strategy builds upon the repeated-sampling-then-voting framework, with a novel twist: incorporating multiple models, even weaker ones, to leverage their complementary strengths that potentially arise from diverse training data and paradigms. By using consistency as a signal, our strategy dynamically switches between models. Theoretical analysis highlights the efficiency and performance advantages of our strategy. Extensive experiments on six datasets demonstrate that our strategy not only outperforms self-consistency and state-of-the-art multi-agent debate approaches, but also significantly reduces inference costs. Additionally, ModelSwitch requires only a few comparable LLMs to achieve optimal performance and can be extended with verification methods, demonstrating the potential of leveraging multiple LLMs in the generation-verification paradigm.

Mixture-of-Experts (MoE) models are crucial for scaling model capacity while controlling inference costs. While integrating MoE into multimodal models like CLIP improves performance, training these models is notoriously challenging and expensive. We propose CLIP-Upcycling (CLIP-UP), an efficient alternative training strategy that converts a pre-trained dense CLIP model into a sparse MoE architecture. Through extensive experimentation with various settings and auxiliary losses, we demonstrate that CLIP-UP significantly reduces training complexity and cost. Remarkably, our sparse CLIP B/16 model, trained with CLIP-UP, outperforms its dense counterpart by 7.2% and 6.6% on COCO and Flickr30k text-to-image Recall@1 benchmarks respectively. It even surpasses the larger CLIP L/14 model on this task while using only 30% of the inference FLOPs. We further demonstrate the generalizability of our training recipe across different scales, establishing sparse upcycling as a practical and scalable approach for building efficient, high-performance CLIP models.

Commercial video generation models have exhibited realistic, high-fidelity results but are still restricted to limited access. One crucial obstacle for large-scale applications is the expensive training and inference cost. In this paper, we argue that videos contain much more redundant information than images, thus can be encoded by very few motion latents based on a content image. Towards this goal, we design an image-conditioned VAE to encode a video to an extremely compressed motion latent space. This magic Reducio charm enables 64x reduction of latents compared to a common 2D VAE, without sacrificing the quality. Training diffusion models on such a compact representation easily allows for generating 1K resolution videos. We then adopt a two-stage video generation paradigm, which performs text-to-image and text-image-to-video sequentially. Extensive experiments show that our Reducio-DiT achieves strong performance in evaluation, though trained with limited GPU resources. More importantly, our method significantly boost the efficiency of video LDMs both in training and inference. We train Reducio-DiT in around 3.2K training hours in total and generate a 16-frame 1024*1024 video clip within 15.5 seconds on a single A100 GPU. Code released at https://github.com/microsoft/Reducio-VAE .

With the rising popularity of Transformer-based large language models (LLMs), reducing their high inference costs has become a significant research focus. One effective approach is to compress the long input contexts. Existing methods typically leverage the self-attention mechanism of the LLM itself for context compression. While these methods have achieved notable results, the compression process still involves quadratic time complexity, which limits their applicability. To mitigate this limitation, we propose the In-Context Former (IC-Former). Unlike previous methods, IC-Former does not depend on the target LLMs. Instead, it leverages the cross-attention mechanism and a small number of learnable digest tokens to directly condense information from the contextual word embeddings. This approach significantly reduces inference time, which achieves linear growth in time complexity within the compression range. Experimental results indicate that our method requires only 1/32 of the floating-point operations of the baseline during compression and improves processing speed by 68 to 112 times while achieving over 90% of the baseline performance on evaluation metrics. Overall, our model effectively reduces compression costs and makes real-time compression scenarios feasible.

Large Language Models (LLMs) have demonstrated impressive performance on multimodal tasks, without any multimodal finetuning. They are the building block for Large Multimodal Models, yet, we still lack a proper understanding of their success. In this work, we expose frozen LLMs to image, video, audio and text inputs and analyse their internal representation aiming to understand their generalization beyond textual inputs. Findings. Perceptual tokens (1) are easily distinguishable from textual ones inside LLMs, with significantly different representations, and complete translation to textual tokens does not exist. Yet, (2) both perceptual and textual tokens activate similar LLM weights. Despite being different, (3) perceptual and textual tokens are implicitly aligned inside LLMs, we call this the implicit multimodal alignment (IMA), and argue that this is linked to architectural design, helping LLMs to generalize. This provide more evidence to believe that the generalization of LLMs to multimodal inputs is mainly due to their architecture. Implications. (1) We find a positive correlation between the implicit alignment score and the task performance, suggesting that this could act as a proxy metric for model evaluation and selection. (2) A negative correlation exists regarding hallucinations, revealing that this problem is mainly due to misalignment between the internal perceptual and textual representations. (3) Perceptual tokens change slightly throughout the model, thus, we propose different approaches to skip computations (e.g. in FFN layers), and significantly reduce the inference cost. (4) Due to the slowly changing embeddings across layers, and the high overlap between textual and multimodal activated weights, we compress LLMs by keeping only 1 subnetwork that works well across a wide range of multimodal tasks. Paper code: https://github.com/mshukor/ima-lmms.

ASR models often suffer from a long-form deletion problem where the model predicts sequential blanks instead of words when transcribing a lengthy audio (in the order of minutes or hours). From the perspective of a user or downstream system consuming the ASR results, this behavior can be perceived as the model "being stuck", and potentially make the product hard to use. One of the culprits for long-form deletion is training-test data mismatch, which can happen even when the model is trained on diverse and large-scale data collected from multiple application domains. In this work, we introduce a novel technique to simultaneously model different groups of speakers in the audio along with the standard transcript tokens. Speakers are grouped as primary and non-primary, which connects the application domains and significantly alleviates the long-form deletion problem. This improved model neither needs any additional training data nor incurs additional training or inference cost.

The energy consumption of deep learning models is increasing at a breathtaking rate, which raises concerns due to potential negative effects on carbon neutrality in the context of global warming and climate change. With the progress of efficient deep learning techniques, e.g., model compression, researchers can obtain efficient models with fewer parameters and smaller latency. However, most of the existing efficient deep learning methods do not explicitly consider energy consumption as a key performance indicator. Furthermore, existing methods mostly focus on the inference costs of the resulting efficient models, but neglect the notable energy consumption throughout the entire life cycle of the algorithm. In this paper, we present the first large-scale energy consumption benchmark for efficient computer vision models, where a new metric is proposed to explicitly evaluate the full-cycle energy consumption under different model usage intensity. The benchmark can provide insights for low carbon emission when selecting efficient deep learning algorithms in different model usage scenarios.

Six-bit quantization (FP6) can effectively reduce the size of large language models (LLMs) and preserve the model quality consistently across varied applications. However, existing systems do not provide Tensor Core support for FP6 quantization and struggle to achieve practical performance improvements during LLM inference. It is challenging to support FP6 quantization on GPUs due to (1) unfriendly memory access of model weights with irregular bit-width and (2) high runtime overhead of weight de-quantization. To address these problems, we propose TC-FPx, the first full-stack GPU kernel design scheme with unified Tensor Core support of float-point weights for various quantization bit-width. We integrate TC-FPx kernel into an existing inference system, providing new end-to-end support (called FP6-LLM) for quantized LLM inference, where better trade-offs between inference cost and model quality are achieved. Experiments show that FP6-LLM enables the inference of LLaMA-70b using only a single GPU, achieving 1.69x-2.65x higher normalized inference throughput than the FP16 baseline. The source code will be publicly available soon.

In recent years, the advent of foundation models (FM) for digital pathology has relied heavily on scaling the pre-training datasets and the model size, yielding large and powerful models. While it resulted in improving the performance on diverse downstream tasks, it also introduced increased computational cost and inference time. In this work, we explore the distillation of a large foundation model into a smaller one, reducing the number of parameters by several orders of magnitude. Leveraging distillation techniques, our distilled model, H0-mini, achieves nearly comparable performance to large FMs at a significantly reduced inference cost. It is evaluated on several public benchmarks, achieving 3rd place on the HEST benchmark and 5th place on the EVA benchmark. Additionally, a robustness analysis conducted on the PLISM dataset demonstrates that our distilled model reaches excellent robustness to variations in staining and scanning conditions, significantly outperforming other state-of-the art models. This opens new perspectives to design lightweight and robust models for digital pathology, without compromising on performance.

Best-of-N (BoN) is a popular and effective algorithm for aligning language models to human preferences. The algorithm works as follows: at inference time, N samples are drawn from the language model, and the sample with the highest reward, as judged by a reward model, is returned as the output. Despite its effectiveness, BoN is computationally expensive; it reduces sampling throughput by a factor of N. To make BoN more efficient at inference time, one strategy is to fine-tune the language model to mimic what BoN does during inference. To achieve this, we derive the distribution induced by the BoN algorithm. We then propose to fine-tune the language model to minimize backward KL divergence to the BoN distribution. Our approach is analogous to mean-field variational inference and, thus, we term it variational BoN (vBoN). To the extent this fine-tuning is successful and we end up with a good approximation, we have reduced the inference cost by a factor of N. Our experiments on a controlled generation task suggest that while variational BoN is not as effective as BoN in aligning language models, it is close to BoN performance as vBoN appears more often on the Pareto frontier of reward and KL divergence compared to models trained with KL-constrained RL objective.

Knowledge distillation (KD) is a common approach to compress a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, in the context of autoregressive language models (LMs), we empirically find that larger teacher LMs might dramatically result in a poorer student. In response to this problem, we conduct a series of analyses and reveal that different tokens have different teaching modes, neglecting which will lead to performance degradation. Motivated by this, we propose a simple yet effective adaptive teaching approach (ATKD) to improve the KD. The core of ATKD is to reduce rote learning and make teaching more diverse and flexible. Extensive experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains (up to +3.04% average score) across all model types and sizes. More encouragingly, ATKD can improve the student model generalization effectively.

We introduce Mistral 7B v0.1, a 7-billion-parameter language model engineered for superior performance and efficiency. Mistral 7B outperforms Llama 2 13B across all evaluated benchmarks, and Llama 1 34B in reasoning, mathematics, and code generation. Our model leverages grouped-query attention (GQA) for faster inference, coupled with sliding window attention (SWA) to effectively handle sequences of arbitrary length with a reduced inference cost. We also provide a model fine-tuned to follow instructions, Mistral 7B -- Instruct, that surpasses the Llama 2 13B -- Chat model both on human and automated benchmarks. Our models are released under the Apache 2.0 license.

Sparse mixture of expert architectures (MoEs) scale model capacity without large increases in training or inference costs. Despite their success, MoEs suffer from a number of issues: training instability, token dropping, inability to scale the number of experts, or ineffective finetuning. In this work, we proposeSoft MoE, a fully-differentiable sparse Transformer that addresses these challenges, while maintaining the benefits of MoEs. Soft MoE performs an implicit soft assignment by passing different weighted combinations of all input tokens to each expert. As in other MoE works, experts in Soft MoE only process a subset of the (combined) tokens, enabling larger model capacity at lower inference cost. In the context of visual recognition, Soft MoE greatly outperforms standard Transformers (ViTs) and popular MoE variants (Tokens Choice and Experts Choice). For example, Soft MoE-Base/16 requires 10.5x lower inference cost (5.7x lower wall-clock time) than ViT-Huge/14 while matching its performance after similar training. Soft MoE also scales well: Soft MoE Huge/14 with 128 experts in 16 MoE layers has over 40x more parameters than ViT Huge/14, while inference time cost grows by only 2%, and it performs substantially better.

Large Language Models (LLMs) have achieved state-of-the-art performance across various language tasks but pose challenges for practical deployment due to their substantial memory requirements. Furthermore, the latest generative models suffer from high inference costs caused by the memory bandwidth bottleneck in the auto-regressive decoding process. To address these issues, we propose an efficient weight-only quantization method that reduces memory consumption and accelerates inference for LLMs. To ensure minimal quality degradation, we introduce a simple and effective heuristic approach that utilizes only the model weights of a pre-trained model. This approach is applicable to both Mixture-of-Experts (MoE) and dense models without requiring additional fine-tuning. To demonstrate the effectiveness of our proposed method, we first analyze the challenges and issues associated with LLM quantization. Subsequently, we present our heuristic approach, which adaptively finds the granularity of quantization, effectively addressing these problems. Furthermore, we implement highly efficient GPU GEMMs that perform on-the-fly matrix multiplication and dequantization, supporting the multiplication of fp16 or bf16 activations with int8 or int4 weights. We evaluate our approach on large-scale open source models such as OPT-175B and internal MoE models, showcasing minimal accuracy loss while achieving up to 3.65 times higher throughput on the same number of GPUs.

Large language models (LLMs) exhibit strong capabilities as decision-making agents by interleaving reasoning and actions, as seen in ReAct-style frameworks. Yet, their practical deployment is constrained by high inference costs and large model sizes. We propose Structured Agent Distillation, a framework that compresses large LLM-based agents into smaller student models while preserving both reasoning fidelity and action consistency. Unlike standard token-level distillation, our method segments trajectories into {[REASON]} and {[ACT]} spans, applying segment-specific losses to align each component with the teacher's behavior. This structure-aware supervision enables compact agents to better replicate the teacher's decision process. Experiments on ALFWorld, HotPotQA-ReAct, and WebShop show that our approach consistently outperforms token-level and imitation learning baselines, achieving significant compression with minimal performance drop. Scaling and ablation results further highlight the importance of span-level alignment for efficient and deployable agents.

We present Neural Attention Search (NAtS), a framework that automatically evaluates the importance of each token within a sequence and determines if the corresponding token can be dropped after several steps. This approach can efficiently reduce the KV cache sizes required by transformer-based models during inference and thus reduce inference costs. In this paper, we design a search space that contains three token types: (i) Global Tokens will be preserved and queried by all the following tokens. (ii) Local Tokens survive until the next global token appears. (iii) Sliding Window Tokens have an impact on the inference of a fixed size of the next following tokens. Similar to the One-Shot Neural Architecture Search approach, this token-type information can be learned jointly with the architecture weights via a learnable attention mask. Experiments on both training a new transformer from scratch and fine-tuning existing large language models show that NAtS can efficiently reduce the KV cache size required for the models while maintaining the models' performance.

We present Generalized LoRA (GLoRA), an advanced approach for universal parameter-efficient fine-tuning tasks. Enhancing Low-Rank Adaptation (LoRA), GLoRA employs a generalized prompt module to optimize pre-trained model weights and adjust intermediate activations, providing more flexibility and capability across diverse tasks and datasets. Moreover, GLoRA facilitates efficient parameter adaptation by employing a scalable, modular, layer-wise structure search that learns individual adapter of each layer. Originating from a unified mathematical formulation, GLoRA exhibits strong transfer learning, few-shot learning and domain generalization abilities, as it adjusts to new tasks through additional dimensions on weights and activations. Comprehensive experiments demonstrate that GLoRA outperforms all previous methods in natural, specialized, and structured benchmarks, achieving superior accuracy with fewer parameters and computations on various datasets. Furthermore, our structural re-parameterization design ensures that GLoRA incurs no extra inference cost, rendering it a practical solution for resource-limited applications. Code is available at: https://github.com/Arnav0400/ViT-Slim/tree/master/GLoRA.

Fill-in-the-Middle (FIM) has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm, which reorders original training sequences and then performs regular next-token prediction (NTP), often leads to models struggling to generate content that aligns smoothly with the surrounding context. Crucially, while existing works rely on rule-based post-processing to circumvent this weakness, such methods are not practically usable in open-domain code completion tasks as they depend on restrictive, dataset-specific assumptions (e.g., generating the same number of lines as in the ground truth). Moreover, model performance on FIM tasks deteriorates significantly without these unrealistic assumptions. We hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens (i.e., horizon length) at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different models and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level, and without resorting to unrealistic post-processing methods. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP only incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.

High-fidelity 3D object synthesis remains significantly more challenging than 2D image generation due to the unstructured nature of mesh data and the cubic complexity of dense volumetric grids. Existing two-stage pipelines-compressing meshes with a VAE (using either 2D or 3D supervision), followed by latent diffusion sampling-often suffer from severe detail loss caused by inefficient representations and modality mismatches introduced in VAE. We introduce Sparc3D, a unified framework that combines a sparse deformable marching cubes representation Sparcubes with a novel encoder Sparconv-VAE. Sparcubes converts raw meshes into high-resolution (1024^3) surfaces with arbitrary topology by scattering signed distance and deformation fields onto a sparse cube, allowing differentiable optimization. Sparconv-VAE is the first modality-consistent variational autoencoder built entirely upon sparse convolutional networks, enabling efficient and near-lossless 3D reconstruction suitable for high-resolution generative modeling through latent diffusion. Sparc3D achieves state-of-the-art reconstruction fidelity on challenging inputs, including open surfaces, disconnected components, and intricate geometry. It preserves fine-grained shape details, reduces training and inference cost, and integrates naturally with latent diffusion models for scalable, high-resolution 3D generation.

This paper identifies significant redundancy in the query-key interactions within self-attention mechanisms of diffusion transformer models, particularly during the early stages of denoising diffusion steps. In response to this observation, we present a novel diffusion transformer framework incorporating an additional set of mediator tokens to engage with queries and keys separately. By modulating the number of mediator tokens during the denoising generation phases, our model initiates the denoising process with a precise, non-ambiguous stage and gradually transitions to a phase enriched with detail. Concurrently, integrating mediator tokens simplifies the attention module's complexity to a linear scale, enhancing the efficiency of global attention processes. Additionally, we propose a time-step dynamic mediator token adjustment mechanism that further decreases the required computational FLOPs for generation, simultaneously facilitating the generation of high-quality images within the constraints of varied inference budgets. Extensive experiments demonstrate that the proposed method can improve the generated image quality while also reducing the inference cost of diffusion transformers. When integrated with the recent work SiT, our method achieves a state-of-the-art FID score of 2.01. The source code is available at https://github.com/LeapLabTHU/Attention-Mediators.

We present MAGI-1, a world model that generates videos by autoregressively predicting a sequence of video chunks, defined as fixed-length segments of consecutive frames. Trained to denoise per-chunk noise that increases monotonically over time, MAGI-1 enables causal temporal modeling and naturally supports streaming generation. It achieves strong performance on image-to-video (I2V) tasks conditioned on text instructions, providing high temporal consistency and scalability, which are made possible by several algorithmic innovations and a dedicated infrastructure stack. MAGI-1 facilitates controllable generation via chunk-wise prompting and supports real-time, memory-efficient deployment by maintaining constant peak inference cost, regardless of video length. The largest variant of MAGI-1 comprises 24 billion parameters and supports context lengths of up to 4 million tokens, demonstrating the scalability and robustness of our approach. The code and models are available at https://github.com/SandAI-org/MAGI-1 and https://github.com/SandAI-org/MagiAttention. The product can be accessed at https://sand.ai.

We introduce sub-sentence encoder, a contrastively-learned contextual embedding model for fine-grained semantic representation of text. In contrast to the standard practice with sentence embeddings, where the meaning of an entire sequence of text is encoded into a fixed-length vector, the sub-sentence encoder learns to produce distinct contextual embeddings corresponding to different atomic propositions, i.e. atomic units of meaning expressed within a text sequence. The sub-sentence embeddings are contrastively learned to recognize (inferred) semantic equivalence between propositions across different text sequences. Our experiments show the effectiveness of sub-sentence encoders in applications, such as retrieving supporting facts for fine-grained text attribution or recognizing the conditional semantic similarity between texts. In practice, we demonstrate that sub-sentence encoders keep the same level of inference cost and space complexity compared to sentence encoders.

We present a latent diffusion model for fast feed-forward 3D scene generation. Given one or more images, our model Bolt3D directly samples a 3D scene representation in less than seven seconds on a single GPU. We achieve this by leveraging powerful and scalable existing 2D diffusion network architectures to produce consistent high-fidelity 3D scene representations. To train this model, we create a large-scale multiview-consistent dataset of 3D geometry and appearance by applying state-of-the-art dense 3D reconstruction techniques to existing multiview image datasets. Compared to prior multiview generative models that require per-scene optimization for 3D reconstruction, Bolt3D reduces the inference cost by a factor of up to 300 times.

Vision Language Models (VLMs) pretrained on Internet-scale vision-language data have demonstrated the potential to transfer their knowledge to robotic learning. However, the existing paradigm encounters three critical challenges: (1) expensive inference cost resulting from large-scale model parameters, (2) frequent domain shifts caused by mismatched data modalities, and (3) limited capacity to handle past or future experiences. In this work, we propose LiteVLP, a lightweight, memory-based, and general-purpose vision-language policy generation model. LiteVLP is built upon a pre-trained 1B-parameter VLM and fine-tuned on a tiny-scale and conversation-style robotic dataset. Through extensive experiments, we demonstrate that LiteVLP outperforms state-of-the-art vision-language policy on VIMA-Bench, with minimal training time. Furthermore, LiteVLP exhibits superior inference speed while maintaining exceptional high accuracy. In long-horizon manipulation tasks, LiteVLP also shows remarkable memory ability, outperforming the best-performing baseline model by 18.8%. These results highlight LiteVLP as a promising model to integrating the intelligence of VLMs into robotic learning.

Recent advances in diffusion-based lip-syncing generative models have demonstrated their ability to produce highly synchronized talking face videos for visual dubbing. Although these models excel at lip synchronization, they often struggle to maintain fine-grained control over facial details in generated images. In this work, we identify "lip averaging" phenomenon where the model fails to preserve subtle facial details when dubbing unseen in-the-wild videos. This issue arises because the commonly used UNet backbone primarily integrates audio features into visual representations in the latent space via cross-attention mechanisms and multi-scale fusion, but it struggles to retain fine-grained lip details in the generated faces. To address this issue, we propose UnAvgLip, which extracts identity embeddings from reference videos to generate highly faithful facial sequences while maintaining accurate lip synchronization. Specifically, our method comprises two primary components: (1) an Identity Perceiver module that encodes facial embeddings to align with conditioned audio features; and (2) an ID-CrossAttn module that injects facial embeddings into the generation process, enhancing model's capability of identity retention. Extensive experiments demonstrate that, at a modest training and inference cost, UnAvgLip effectively mitigates the "averaging" phenomenon in lip inpainting, significantly preserving unique facial characteristics while maintaining precise lip synchronization. Compared with the original approach, our method demonstrates significant improvements of 5% on the identity consistency metric and 2% on the SSIM metric across two benchmark datasets (HDTF and LRW).

Modern recurrent layers are emerging as a promising path toward edge deployment of foundation models, especially in the context of large language models (LLMs). Compressing the whole input sequence in a finite-dimensional representation enables recurrent layers to model long-range dependencies while maintaining a constant inference cost for each token and a fixed memory requirement. However, the practical deployment of LLMs in resource-limited environments often requires further model compression, such as quantization and pruning. While these techniques are well-established for attention-based models, their effects on recurrent layers remain underexplored. In this preliminary work, we focus on post-training quantization for recurrent LLMs and show that Mamba models exhibit the same pattern of outlier channels observed in attention-based LLMs. We show that the reason for the difficulty of quantizing SSMs is caused by activation outliers, similar to those observed in transformer-based LLMs. We report baseline results for post-training quantization of Mamba that do not take into account the activation outliers and suggest first steps for outlier-aware quantization.

Instruction finetuning on a variety of image-text instruction data is the key to obtaining a versatile Multimodal Large Language Model (MLLM), and different configurations of the instruction data can lead to finetuned models with different capabilities. However, we have discovered that data conflicts are inevitable when mixing instruction data from distinct domains, which can result in performance drops for tasks of a specific domain. To address this issue, we propose to apply an efficient Mixture of Experts (MoE) design, which is a sparse Mixture of LoRA Experts (MoLE) for instruction finetuning MLLMs. Within the Transformer layers, we extend the popular Low-Rank Adaption (LoRA) method by creating a set of LoRA experts specifically for the MLP layer, and route each token to the top-1 expert based on a routing function, allowing adaptive choices for tokens from different domains. Since the LoRA experts are sparsely activated, the training and inference cost are kept roughly constant compared to the original LoRA method. By replacing the plain-LoRA of LLaVA-1.5 with our MoE design, our final model is named LLaVA-MoLE. Extensive experiments proved that LLaVA-MoLE effectively mitigates the data conflict issue when mixing multiple distinct instruction datasets with various configurations, and achieves consistent performance gains over the strong plain-LoRA baselines. Most importantly, on the mixed datasets, LLaVA-MoLE can even outperform the plain-LoRA baseline trained with twice the samples.

Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, ensembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of ensembling to that of a single model. TMC improves accuracy by 4.2% compared to ensembling non-linearly fine-tuned models at a 2.5x to 10x reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting -- task-incremental, class-incremental, and data-incremental -- on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings.

Scaling laws have been recently employed to derive compute-optimal model size (number of parameters) for a given compute duration. We advance and refine such methods to infer compute-optimal model shapes, such as width and depth, and successfully implement this in vision transformers. Our shape-optimized vision transformer, SoViT, achieves results competitive with models that exceed twice its size, despite being pre-trained with an equivalent amount of compute. For example, SoViT-400m/14 achieves 90.3% fine-tuning accuracy on ILSRCV2012, surpassing the much larger ViT-g/14 and approaching ViT-G/14 under identical settings, with also less than half the inference cost. We conduct a thorough evaluation across multiple tasks, such as image classification, captioning, VQA and zero-shot transfer, demonstrating the effectiveness of our model across a broad range of domains and identifying limitations. Overall, our findings challenge the prevailing approach of blindly scaling up vision models and pave a path for a more informed scaling.

Policy constraint methods to offline reinforcement learning (RL) typically utilize parameterization or regularization that constrains the policy to perform actions within the support set of the behavior policy. The elaborative designs of parameterization methods usually intrude into the policy networks, which may bring extra inference cost and cannot take full advantage of well-established online methods. Regularization methods reduce the divergence between the learned policy and the behavior policy, which may mismatch the inherent density-based definition of support set thereby failing to avoid the out-of-distribution actions effectively. This paper presents Supported Policy OpTimization (SPOT), which is directly derived from the theoretical formalization of the density-based support constraint. SPOT adopts a VAE-based density estimator to explicitly model the support set of behavior policy and presents a simple but effective density-based regularization term, which can be plugged non-intrusively into off-the-shelf off-policy RL algorithms. SPOT achieves the state-of-the-art performance on standard benchmarks for offline RL. Benefiting from the pluggable design, offline pretrained models from SPOT can also be applied to perform online fine-tuning seamlessly.

Chain-of-thought (CoT) via prompting is the de facto method for eliciting reasoning capabilities from large language models (LLMs). But for what kinds of tasks is this extra ``thinking'' really helpful? To analyze this, we conducted a quantitative meta-analysis covering over 100 papers using CoT and ran our own evaluations of 20 datasets across 14 models. Our results show that CoT gives strong performance benefits primarily on tasks involving math or logic, with much smaller gains on other types of tasks. On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's response contains an equals sign, indicating symbolic operations and reasoning. Following this finding, we analyze the behavior of CoT on these problems by separating planning and execution and comparing against tool-augmented LLMs. Much of CoT's gain comes from improving symbolic execution, but it underperforms relative to using a symbolic solver. Our results indicate that CoT can be applied selectively, maintaining performance while saving inference costs. Furthermore, they suggest a need to move beyond prompt-based CoT to new paradigms that better leverage intermediate computation across the whole range of LLM applications.

Recent advancements in text-to-image generative systems have been largely driven by diffusion models. However, single-stage text-to-image diffusion models still face challenges, in terms of computational efficiency and the refinement of image details. To tackle the issue, we propose CogView3, an innovative cascaded framework that enhances the performance of text-to-image diffusion. CogView3 is the first model implementing relay diffusion in the realm of text-to-image generation, executing the task by first creating low-resolution images and subsequently applying relay-based super-resolution. This methodology not only results in competitive text-to-image outputs but also greatly reduces both training and inference costs. Our experimental results demonstrate that CogView3 outperforms SDXL, the current state-of-the-art open-source text-to-image diffusion model, by 77.0\% in human evaluations, all while requiring only about 1/2 of the inference time. The distilled variant of CogView3 achieves comparable performance while only utilizing 1/10 of the inference time by SDXL.

Recent methods have demonstrated that Large Language Models (LLMs) can solve reasoning tasks better when they are encouraged to solve subtasks of the main task first. In this paper we devise a similar strategy that breaks down reasoning tasks into a problem decomposition phase and a problem solving phase and show that the strategy is able to outperform a single stage solution. Further, we hypothesize that the decomposition should be easier to distill into a smaller model compared to the problem solving because the latter requires large amounts of domain knowledge while the former only requires learning general problem solving strategies. We propose methods to distill these two capabilities and evaluate their impact on reasoning outcomes and inference cost. We find that we can distill the problem decomposition phase and at the same time achieve good generalization across tasks, datasets, and models. However, it is harder to distill the problem solving capability without losing performance and the resulting distilled model struggles with generalization. These results indicate that by using smaller, distilled problem decomposition models in combination with problem solving LLMs we can achieve reasoning with cost-efficient inference and local adaptation.

Recent research has leveraged large language model multi-agent systems for complex problem-solving while trying to reduce the manual effort required to build them, driving the development of automated agent workflow optimization methods. However, existing methods remain inflexible due to representational limitations, a lack of adaptability, and poor scalability when relying on discrete optimization techniques. We address these challenges with ScoreFlow, a simple yet high-performance framework that leverages efficient gradient-based optimization in a continuous space. ScoreFlow incorporates Score-DPO, a novel variant of the direct preference optimization method that accounts for quantitative feedback. Across six benchmarks spanning question answering, coding, and mathematical reasoning, ScoreFlow achieves an 8.2% improvement over existing baselines. Moreover, it empowers smaller models to outperform larger ones with lower inference costs. Project: https://github.com/Gen-Verse/ScoreFlow

Personalized image generation requires text-to-image generative models that capture the core features of a reference subject to allow for controlled generation across different contexts. Existing methods face challenges due to complex training requirements, high inference costs, limited flexibility, or a combination of these issues. In this paper, we introduce DreamCache, a scalable approach for efficient and high-quality personalized image generation. By caching a small number of reference image features from a subset of layers and a single timestep of the pretrained diffusion denoiser, DreamCache enables dynamic modulation of the generated image features through lightweight, trained conditioning adapters. DreamCache achieves state-of-the-art image and text alignment, utilizing an order of magnitude fewer extra parameters, and is both more computationally effective and versatile than existing models.

Large language models (LLMs) are powerful tools for content moderation, but their inference costs and latency make them prohibitive for casual use on large datasets, such as the Google Ads repository. This study proposes a method for scaling up LLM reviews for content moderation in Google Ads. First, we use heuristics to select candidates via filtering and duplicate removal, and create clusters of ads for which we select one representative ad per cluster. We then use LLMs to review only the representative ads. Finally, we propagate the LLM decisions for the representative ads back to their clusters. This method reduces the number of reviews by more than 3 orders of magnitude while achieving a 2x recall compared to a baseline non-LLM model. The success of this approach is a strong function of the representations used in clustering and label propagation; we found that cross-modal similarity representations yield better results than uni-modal representations.

Mixture of Experts (MoE) has become a key ingredient for scaling large foundation models while keeping inference costs steady. We show that expert routing strategies that have cross-batch dependencies are vulnerable to attacks. Malicious queries can be sent to a model and can affect a model's output on other benign queries if they are grouped in the same batch. We demonstrate this via a proof-of-concept attack in a toy experimental setting.

Achieving precise alignment between textual instructions and generated images in text-to-image generation is a significant challenge, particularly in rendering written text within images. Sate-of-the-art models like Stable Diffusion 3 (SD3), Flux, and AuraFlow still struggle with accurate text depiction, resulting in misspelled or inconsistent text. We introduce a training-free method with minimal computational overhead that significantly enhances text rendering quality. Specifically, we introduce an overshooting sampler for pretrained rectified flow (RF) models, by alternating between over-simulating the learned ordinary differential equation (ODE) and reintroducing noise. Compared to the Euler sampler, the overshooting sampler effectively introduces an extra Langevin dynamics term that can help correct the compounding error from successive Euler steps and therefore improve the text rendering. However, when the overshooting strength is high, we observe over-smoothing artifacts on the generated images. To address this issue, we propose an Attention Modulated Overshooting sampler (AMO), which adaptively controls the strength of overshooting for each image patch according to their attention score with the text content. AMO demonstrates a 32.3% and 35.9% improvement in text rendering accuracy on SD3 and Flux without compromising overall image quality or increasing inference cost.

Emerging large reasoning models (LRMs), such as DeepSeek-R1 models, leverage long chain-of-thought (CoT) reasoning to generate structured intermediate steps, enhancing their reasoning capabilities. However, long CoT does not inherently guarantee safe outputs, potentially leading to harmful consequences such as the introduction of security vulnerabilities in code or the spread of misinformation. Current research on large language model (LLM) safety usually focuses on short-answer responses, overlooking the long CoT style outputs of LRMs. To bridge this gap, we conduct a systematic study of LRM safety. First, we investigate safety evaluators calibrated against human annotations. Using our newly developed metrics, we thoroughly assess the safety of 12 state-of-the-art LRMs on StrongReject and WildJailbreak datasets. Our results show that LRMs are not safe compared to their reasoning advance. Further, we perform a fine-grained analysis of the reasoning trace and final answer. We find that three decoding strategies-ZeroThink, LessThink, and MoreThink-can improve model safety without additional training. However, these strategies either use constrained reasoning traces or incur high inference costs. To better strengthen LRM safety, we introduce SafeChain, the first-of-its-kind safety training dataset in CoT style. We fine-tune two LRMs with SafeChain, showing that it not only enhances model safety but also preserves performance across 6 reasoning benchmarks.

Recent advances in text-to-image diffusion models have enabled the photorealistic generation of images from text prompts. Despite the great progress, existing models still struggle to generate compositional multi-concept images naturally, limiting their ability to visualize human imagination. While several recent works have attempted to address this issue, they either introduce additional training or adopt guidance at inference time. In this work, we consider a more ambitious goal: natural multi-concept generation using a pre-trained diffusion model, and with almost no extra cost. To achieve this goal, we identify the limitations in the text embeddings used for the pre-trained text-to-image diffusion models. Specifically, we observe concept dominance and non-localized contribution that severely degrade multi-concept generation performance. We further design a minimal low-cost solution that overcomes the above issues by tweaking (not re-training) the text embeddings for more realistic multi-concept text-to-image generation. Our Correction by Similarities method tweaks the embedding of concepts by collecting semantic features from most similar tokens to localize the contribution. To avoid mixing features of concepts, we also apply Cross-Token Non-Maximum Suppression, which excludes the overlap of contributions from different concepts. Experiments show that our approach outperforms previous methods in text-to-image, image manipulation, and personalization tasks, despite not introducing additional training or inference costs to the diffusion steps.

Structural re-parameterization is a general training scheme for Convolutional Neural Networks (CNNs), which achieves performance improvement without increasing inference cost. As Vision Transformers (ViTs) are gradually surpassing CNNs in various visual tasks, one may question: if a training scheme specifically for ViTs exists that can also achieve performance improvement without increasing inference cost? Recently, Mixture-of-Experts (MoE) has attracted increasing attention, as it can efficiently scale up the capacity of Transformers at a fixed cost through sparsely activated experts. Considering that MoE can also be viewed as a multi-branch structure, can we utilize MoE to implement a ViT training scheme similar to structural re-parameterization? In this paper, we affirmatively answer these questions, with a new general training strategy for ViTs. Specifically, we decouple the training and inference phases of ViTs. During training, we replace some Feed-Forward Networks (FFNs) of the ViT with specially designed, more efficient MoEs that assign tokens to experts by random uniform partition, and perform Experts Weights Averaging (EWA) on these MoEs at the end of each iteration. After training, we convert each MoE into an FFN by averaging the experts, transforming the model back into original ViT for inference. We further provide a theoretical analysis to show why and how it works. Comprehensive experiments across various 2D and 3D visual tasks, ViT architectures, and datasets validate the effectiveness and generalizability of the proposed training scheme. Besides, our training scheme can also be applied to improve performance when fine-tuning ViTs. Lastly, but equally important, the proposed EWA technique can significantly improve the effectiveness of naive MoE in various 2D visual small datasets and 3D visual tasks.

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by retrieving relevant documents from external sources and incorporating them into the context. While it improves reliability by providing factual texts, it significantly increases inference costs as context length grows and introduces challenging issue of RAG hallucination, primarily caused by the lack of corresponding parametric knowledge in LLMs. An efficient solution is to enhance the knowledge of LLMs at test-time. Parametric RAG (PRAG) addresses this by embedding document into LLMs parameters to perform test-time knowledge enhancement, effectively reducing inference costs through offline training. However, its high training and storage costs, along with limited generalization ability, significantly restrict its practical adoption. To address these challenges, we propose Dynamic Parametric RAG (DyPRAG), a novel framework that leverages a lightweight parameter translator model to efficiently convert documents into parametric knowledge. DyPRAG not only reduces inference, training, and storage costs but also dynamically generates parametric knowledge, seamlessly enhancing the knowledge of LLMs and resolving knowledge conflicts in a plug-and-play manner at test-time. Extensive experiments on multiple datasets demonstrate the effectiveness and generalization capabilities of DyPRAG, offering a powerful and practical RAG paradigm which enables superior knowledge fusion and mitigates RAG hallucination in real-world applications. Our code is available at https://github.com/Trae1ounG/DyPRAG.

Achieving optimal performance of video diffusion transformers within given data and compute budget is crucial due to their high training costs. This necessitates precisely determining the optimal model size and training hyperparameters before large-scale training. While scaling laws are employed in language models to predict performance, their existence and accurate derivation in visual generation models remain underexplored. In this paper, we systematically analyze scaling laws for video diffusion transformers and confirm their presence. Moreover, we discover that, unlike language models, video diffusion models are more sensitive to learning rate and batch size, two hyperparameters often not precisely modeled. To address this, we propose a new scaling law that predicts optimal hyperparameters for any model size and compute budget. Under these optimal settings, we achieve comparable performance and reduce inference costs by 40.1% compared to conventional scaling methods, within a compute budget of 1e10 TFlops. Furthermore, we establish a more generalized and precise relationship among validation loss, any model size, and compute budget. This enables performance prediction for non-optimal model sizes, which may also be appealed under practical inference cost constraints, achieving a better trade-off.

Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top K routed experts in an ``addition'' manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a ``multiplication'' manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.

Scaling the size of a model enhances its capabilities but significantly increases computation complexity. Mixture-of-Experts models (MoE) address the issue by allowing model size to scale up without substantially increasing training or inference costs. Despite their promising results, MoE models encounter several challenges. Primarily, for dynamic routing methods, the dispersion of training tokens across multiple experts can lead to underfitting, particularly for infrequent tokens. Additionally, while fixed routing methods can mitigate that issue, they compromise on the diversity of representations. In this paper, we propose MaskMoE, a method designed to enhance token-level learning by employing a routing masking technique within the Mixture-of-Experts model. MaskMoE is capable of maintaining representation diversity while achieving more comprehensive training. Experimental results demonstrate that our method outperforms previous dominant Mixture-of-Experts models in terms of both perplexity (PPL) and downstream task performance.

Despite the promise of Mixture of Experts (MoE) models in increasing parameter counts of Transformer models while maintaining training and inference costs, their application carries notable drawbacks. The key strategy of these models is to, for each processed token, activate at most a few experts - subsets of an extensive feed-forward layer. But this approach is not without its challenges. The operation of matching experts and tokens is discrete, which makes MoE models prone to issues like training instability and uneven expert utilization. Existing techniques designed to address these concerns, such as auxiliary losses or balance-aware matching, result either in lower model performance or are more difficult to train. In response to these issues, we propose Mixture of Tokens, a fully-differentiable model that retains the benefits of MoE architectures while avoiding the aforementioned difficulties. Rather than routing tokens to experts, this approach mixes tokens from different examples prior to feeding them to experts, enabling the model to learn from all token-expert combinations. Importantly, this mixing can be disabled to avoid mixing of different sequences during inference. Crucially, this method is fully compatible with both masked and causal Large Language Model training and inference.

Large language models have become a vital component in modern NLP, achieving state of the art performance in a variety of tasks. However, they are often inefficient for real-world deployment due to their expensive inference costs. Knowledge distillation is a promising technique to improve their efficiency while retaining most of their effectiveness. In this paper, we reproduce, compare and analyze several representative methods for task-agnostic (general-purpose) distillation of Transformer language models. Our target of study includes Output Distribution (OD) transfer, Hidden State (HS) transfer with various layer mapping strategies, and Multi-Head Attention (MHA) transfer based on MiniLMv2. Through our extensive experiments, we study the effectiveness of each method for various student architectures in both monolingual (English) and multilingual settings. Overall, we show that MHA transfer based on MiniLMv2 is generally the best option for distillation and explain the potential reasons behind its success. Moreover, we show that HS transfer remains as a competitive baseline, especially under a sophisticated layer mapping strategy, while OD transfer consistently lags behind other approaches. Findings from this study helped us deploy efficient yet effective student models for latency-critical applications.

Image recognition has recently witnessed a paradigm shift, where vision-language models are now used to perform few-shot classification based on textual prompts. Among these, the CLIP model has shown remarkable capabilities for zero-shot transfer by matching an image and a custom textual prompt in its latent space. This has paved the way for several works that focus on engineering or learning textual contexts for maximizing CLIP's classification capabilities. In this paper, we follow this trend by learning an ensemble of prompts for image classification. We show that learning diverse and possibly shorter contexts improves considerably and consistently the results rather than relying on a single trainable prompt. In particular, we report better few-shot capabilities with no additional cost at inference time. We demonstrate the capabilities of our approach on 11 different benchmarks.

Facial expression recognition (FER) remains a challenging task due to the ambiguity of expressions. The derived noisy labels significantly harm the performance in real-world scenarios. To address this issue, we present a new FER model named Landmark-Aware Net~(LA-Net), which leverages facial landmarks to mitigate the impact of label noise from two perspectives. Firstly, LA-Net uses landmark information to suppress the uncertainty in expression space and constructs the label distribution of each sample by neighborhood aggregation, which in turn improves the quality of training supervision. Secondly, the model incorporates landmark information into expression representations using the devised expression-landmark contrastive loss. The enhanced expression feature extractor can be less susceptible to label noise. Our method can be integrated with any deep neural network for better training supervision without introducing extra inference costs. We conduct extensive experiments on both in-the-wild datasets and synthetic noisy datasets and demonstrate that LA-Net achieves state-of-the-art performance.

Transformer-based language models (LMs) are powerful and widely-applicable tools, but their usefulness is constrained by a finite context window and the expensive computational cost of processing long text documents. We propose to adapt pre-trained LMs into AutoCompressors. These models are capable of compressing long contexts into compact summary vectors, which are then accessible to the model as soft prompts. Summary vectors are trained with an unsupervised objective, whereby long documents are processed in segments and summary vectors from all previous segments are used in language modeling. We fine-tune OPT models on sequences of up to 30,720 tokens and show that AutoCompressors can utilize long contexts to improve perplexity. We evaluate AutoCompressors on in-context learning by compressing task demonstrations. We find that summary vectors are good substitutes for plain-text demonstrations, increasing accuracy while reducing inference cost. Finally, we explore the benefits of pre-computing summary vectors for large corpora by applying summary vectors to retrieval-augmented language modeling. Overall, AutoCompressors emerge as a simple and inexpensive solution for extending the context window of LMs while speeding up inference over long contexts.

Despite tremendous success in many application scenarios, the training and inference costs of using deep learning are also rapidly increasing over time. The lottery ticket hypothesis (LTH) emerges as a promising framework to leverage a special sparse subnetwork (i.e., winning ticket) instead of a full model for both training and inference, that can lower both costs without sacrificing the performance. The main resource bottleneck of LTH is however the extraordinary cost to find the sparse mask of the winning ticket. That makes the found winning ticket become a valuable asset to the owners, highlighting the necessity of protecting its copyright. Our setting adds a new dimension to the recently soaring interest in protecting against the intellectual property (IP) infringement of deep models and verifying their ownerships, since they take owners' massive/unique resources to develop or train. While existing methods explored encrypted weights or predictions, we investigate a unique way to leverage sparse topological information to perform lottery verification, by developing several graph-based signatures that can be embedded as credentials. By further combining trigger set-based methods, our proposal can work in both white-box and black-box verification scenarios. Through extensive experiments, we demonstrate the effectiveness of lottery verification in diverse models (ResNet-20, ResNet-18, ResNet-50) on CIFAR-10 and CIFAR-100. Specifically, our verification is shown to be robust to removal attacks such as model fine-tuning and pruning, as well as several ambiguity attacks. Our codes are available at https://github.com/VITA-Group/NO-stealing-LTH.

As Large Language Models (LLMs) rapidly advance, we introduce Hunyuan-TurboS, a novel large hybrid Transformer-Mamba Mixture of Experts (MoE) model. It synergistically combines Mamba's long-sequence processing efficiency with Transformer's superior contextual understanding. Hunyuan-TurboS features an adaptive long-short chain-of-thought (CoT) mechanism, dynamically switching between rapid responses for simple queries and deep "thinking" modes for complex problems, optimizing computational resources. Architecturally, this 56B activated (560B total) parameter model employs 128 layers (Mamba2, Attention, FFN) with an innovative AMF/MF block pattern. Faster Mamba2 ensures linear complexity, Grouped-Query Attention minimizes KV cache, and FFNs use an MoE structure. Pre-trained on 16T high-quality tokens, it supports a 256K context length and is the first industry-deployed large-scale Mamba model. Our comprehensive post-training strategy enhances capabilities via Supervised Fine-Tuning (3M instructions), a novel Adaptive Long-short CoT Fusion method, Multi-round Deliberation Learning for iterative improvement, and a two-stage Large-scale Reinforcement Learning process targeting STEM and general instruction-following. Evaluations show strong performance: overall top 7 rank on LMSYS Chatbot Arena with a score of 1356, outperforming leading models like Gemini-2.0-Flash-001 (1352) and o4-mini-2025-04-16 (1345). TurboS also achieves an average of 77.9% across 23 automated benchmarks. Hunyuan-TurboS balances high performance and efficiency, offering substantial capabilities at lower inference costs than many reasoning models, establishing a new paradigm for efficient large-scale pre-trained models.

Real-world RAG applications often encounter long-context input scenarios, where redundant information and noise results in higher inference costs and reduced performance. To address these challenges, we propose LongRefiner, an efficient plug-and-play refiner that leverages the inherent structural characteristics of long documents. LongRefiner employs dual-level query analysis, hierarchical document structuring, and adaptive refinement through multi-task learning on a single foundation model. Experiments on seven QA datasets demonstrate that LongRefiner achieves competitive performance in various scenarios while using 10x fewer computational costs and latency compared to the best baseline. Further analysis validates that LongRefiner is scalable, efficient, and effective, providing practical insights for real-world long-text RAG applications. Our code is available at https://github.com/ignorejjj/LongRefiner.

Sparse Mixture of Experts (SMoE) enables efficient training of large language models by routing input tokens to a select number of experts. However, training SMoE remains challenging due to the issue of representation collapse. Recent studies have focused on improving the router to mitigate this problem, but existing approaches face two key limitations: (1) expert embeddings are significantly smaller than the model's dimension, contributing to representation collapse, and (2) routing each input to the Top-K experts can cause them to learn overly similar features. In this work, we propose a novel approach called Robust Sparse Mixture of Experts via Stochastic Learning (S2MoE), which is a mixture of experts designed to learn from both deterministic and non-deterministic inputs via Learning under Uncertainty. Extensive experiments across various tasks demonstrate that S2MoE achieves performance comparable to other routing methods while reducing computational inference costs by 28%.

Sparse Mixture of Experts (SMoE) improves the efficiency of large language model training by directing input tokens to a subset of experts. Despite its success in generation tasks, its generalization ability remains an open question. In this paper, we demonstrate that current SMoEs, which fall into two categories: (1) Token Choice ;and (2) Expert Choice, struggle with tasks such as the Massive Text Embedding Benchmark (MTEB). By analyzing their mechanism through the lens of competitive learning, our study finds that the Token Choice approach may overly focus on irrelevant experts, while the Expert Choice approach risks discarding important tokens, potentially affecting performance. Motivated by this analysis, we propose Unified Competitive Learning SMoE (USMoE), a novel and efficient framework designed to improve the performance of existing SMoEs in both scenarios: with and without training. Extensive experiments across various tasks show that USMoE achieves up to a 10% improvement over traditional approaches or reduces computational inference costs by 14% while maintaining strong performance.

Large Language Model (LLM)-empowered multi-agent systems extend the cognitive boundaries of individual agents through disciplined collaboration and interaction, while constructing these systems often requires labor-intensive manual designs. Despite the availability of methods to automate the design of agentic workflows, they typically seek to identify a static, complex, one-size-fits-all system, which, however, fails to dynamically allocate inference resources based on the difficulty and domain of each query. To address this challenge, we shift away from the pursuit of a monolithic agentic system, instead optimizing the agentic supernet, a probabilistic and continuous distribution of agentic architectures. We introduce MaAS, an automated framework that samples query-dependent agentic systems from the supernet, delivering high-quality solutions and tailored resource allocation (e.g., LLM calls, tool calls, token cost). Comprehensive evaluation across six benchmarks demonstrates that MaAS (I) requires only 6sim45% of the inference costs of existing handcrafted or automated multi-agent systems, (II) surpasses them by 0.54%sim11.82%, and (III) enjoys superior cross-dataset and cross-LLM-backbone transferability.

The adoption of Foundation Models in resource-constrained environments remains challenging due to their large size and inference costs. A promising way to overcome these limitations is post-training compression, which aims to balance reduced model size against performance degradation. This work presents Any Compression via Iterative Pruning (ACIP), a novel algorithmic approach to determine a compression-performance trade-off from a single stochastic gradient descent run. To ensure parameter efficiency, we use an SVD-reparametrization of linear layers and iteratively prune their singular values with a sparsity-inducing penalty. The resulting pruning order gives rise to a global parameter ranking that allows us to materialize models of any target size. Importantly, the compressed models exhibit strong predictive downstream performance without the need for costly fine-tuning. We evaluate ACIP on a large selection of open-weight LLMs and tasks, and demonstrate state-of-the-art results compared to existing factorisation-based compression methods. We also show that ACIP seamlessly complements common quantization-based compression techniques.

Solving complex real-world tasks requires cycles of actions and observations. This is particularly true in science, where tasks require many cycles of analysis, tool use, and experimentation. Language agents are promising for automating intellectual tasks in science because they can interact with tools via natural language or code. Yet their flexibility creates conceptual and practical challenges for software implementations, since agents may comprise non-standard components such as internal reasoning, planning, tool usage, as well as the inherent stochasticity of temperature-sampled language models. Here, we introduce Aviary, an extensible gymnasium for language agents. We formalize agents as policies solving language-grounded partially observable Markov decision processes, which we term language decision processes. We then implement five environments, including three challenging scientific environments: (1) manipulating DNA constructs for molecular cloning, (2) answering research questions by accessing scientific literature, and (3) engineering protein stability. These environments were selected for their focus on multi-step reasoning and their relevance to contemporary biology research. Finally, with online training and scaling inference-time compute, we show that language agents backed by open-source, non-frontier LLMs can match and exceed both frontier LLM agents and human experts on multiple tasks at up to 100x lower inference cost.

Fine-grained sentiment analysis (FSA) aims to extract and summarize user opinions from vast opinionated text. Recent studies demonstrate that large language models (LLMs) possess exceptional sentiment understanding capabilities. However, directly deploying LLMs for FSA applications incurs high inference costs. Therefore, this paper investigates the distillation of fine-grained sentiment understanding from LLMs into small language models (SLMs). We prompt LLMs to examine and interpret the sentiments of given reviews and then utilize the generated content to pretrain SLMs. Additionally, we develop a comprehensive FSA benchmark to evaluate both SLMs and LLMs. Extensive experiments on this benchmark reveal that: (1) distillation significantly enhances the performance of SLMs in FSA tasks, achieving a 6.00\% improvement in F_1-score, and the distilled model can outperform Llama-2-7b with only 220M parameters; (2) distillation equips SLMs with excellent zero-shot sentiment classification capabilities, enabling them to match or even exceed their teacher models. These results suggest that distillation from LLMs is a highly promising direction for FSA. We will release our code, data, and pretrained model weights at https://github.com/HITSZ-HLT/FSA-Distillation.

In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks directly from the given demonstrations without requiring gradient updates. While recent advances have expanded context windows to accommodate more demonstrations, this approach increases inference costs without necessarily improving performance. To mitigate these issues, We propose StreamAdapter, a novel approach that directly updates model parameters from context at test time, eliminating the need for explicit in-context demonstrations. StreamAdapter employs context mapping and weight absorption mechanisms to dynamically transform ICL demonstrations into parameter updates with minimal additional parameters. By reducing reliance on numerous in-context examples, StreamAdapter significantly reduce inference costs and allows for efficient inference with constant time complexity, regardless of demonstration count. Extensive experiments across diverse tasks and model architectures demonstrate that StreamAdapter achieves comparable or superior adaptation capability to ICL while requiring significantly fewer demonstrations. The superior task adaptation and context encoding capabilities of StreamAdapter on both language understanding and generation tasks provides a new perspective for adapting LLMs at test time using context, allowing for more efficient adaptation across scenarios and more cost-effective inference

Large language models (LLMs) exhibit remarkable reasoning abilities, allowing them to generalize across a wide range of downstream tasks, such as commonsense reasoning or instruction following. However, as LLMs scale, inference costs become increasingly prohibitive, accumulating significantly over their life cycle. This poses the question: Can we compress pre-trained LLMs to meet diverse size and latency requirements? We leverage Neural Architecture Search (NAS) to compress LLMs by pruning structural components, such as attention heads, neurons, and layers, aiming to achieve a Pareto-optimal balance between performance and efficiency. While NAS already achieved promising results on small language models in previous work, in this paper we propose various extensions that allow us to scale to LLMs. Compared to structural pruning baselines, we show that NAS improves performance up to 3.4% on MMLU with an on-device latency speedup.

Being prompted to engage in reasoning has emerged as a core technique for using large language models (LLMs), deploying additional inference-time compute to improve task performance. However, as LLMs increase in both size and adoption, inference costs are correspondingly becoming increasingly burdensome. How, then, might we optimize reasoning's cost-performance tradeoff? This work introduces a novel approach based on computational models of metareasoning used in cognitive science, training LLMs to selectively use intermediate reasoning steps only when necessary. We first develop a reward function that incorporates the Value of Computation by penalizing unnecessary reasoning, then use this reward function with Expert Iteration to train the LLM. Compared to few-shot chain-of-thought prompting and STaR, our method significantly reduces inference costs (20-37\% fewer tokens generated across three models) while maintaining task performance across diverse datasets.

Low-rank adaptation (LoRA) and its variants have recently gained much interest due to their ability to avoid excessive inference costs. However, LoRA still encounters the following challenges: (1) Limitation of low-rank assumption; and (2) Its initialization method may be suboptimal. To this end, we propose PMSS(Pre-trained Matrices Skeleton Selection), which enables high-rank updates with low costs while leveraging semantic and linguistic information inherent in pre-trained weight. It achieves this by selecting skeletons from the pre-trained weight matrix and only learning a small matrix instead. Experiments demonstrate that PMSS outperforms LoRA and other fine-tuning methods across tasks with much less trainable parameters. We demonstrate its effectiveness, especially in handling complex tasks such as DROP benchmark(+3.4%/+5.9% on LLaMA2-7B/13B) and math reasoning(+12.89%/+5.61%/+3.11% on LLaMA2-7B, Mistral-7B and Gemma-7B of GSM8K). The code and model will be released soon.

Small Language Models (SLMs) are attracting attention due to the high computational demands and privacy concerns of Large Language Models (LLMs). Some studies fine-tune SLMs using Chains of Thought (CoT) data distilled from LLMs, aiming to enhance their reasoning ability. Furthermore, Some CoT distillation methods introduce external symbolic knowledge into the generation process to improve the limited knowledge memory, reasoning ability and out-of-domain (OOD) generalization of SLMs. However, the introduction of symbolic knowledge increases computational overhead and introduces potential noise. In this paper, we introduce SKIntern, an innovative approach that empowers SLMs to internalize symbolic knowledge and few-shot examples gradually through a progressive fine-tuning process, guided by a predefined linear decay schedule under curriculum learning. By efficiently internalizing knowledge, SKIntern reduces computational overhead and speeds up the reasoning process by focusing solely on the question during inference. It outperforms state-of-the-art baselines by over 5\%, while reducing inference costs (measured in FLOPs) by up to 4times across a wide range of SLMs in both in-domain (ID) and out-of-domain (OOD) tasks. Our code will be available at https://github.com/Xnhyacinth/SKIntern.

Most text-video retrieval methods utilize the text-image pre-trained models like CLIP as a backbone. These methods process each sampled frame independently by the image encoder, resulting in high computational overhead and limiting practical deployment. Addressing this, we focus on efficient text-video retrieval by tackling two key challenges: 1. From the perspective of trainable parameters, current parameter-efficient fine-tuning methods incur high inference costs; 2. From the perspective of model complexity, current token compression methods are mainly designed for images to reduce spatial redundancy but overlook temporal redundancy in consecutive frames of a video. To tackle these challenges, we propose Temporal Token Merging (TempMe), a parameter-efficient and training-inference efficient text-video retrieval architecture that minimizes trainable parameters and model complexity. Specifically, we introduce a progressive multi-granularity framework. By gradually combining neighboring clips, we reduce spatio-temporal redundancy and enhance temporal modeling across different frames, leading to improved efficiency and performance. Extensive experiments validate the superiority of our TempMe. Compared to previous parameter-efficient text-video retrieval methods, TempMe achieves superior performance with just 0.50M trainable parameters. It significantly reduces output tokens by 95% and GFLOPs by 51%, while achieving a 1.8X speedup and a 4.4% R-Sum improvement. With full fine-tuning, TempMe achieves a significant 7.9% R-Sum improvement, trains 1.57X faster, and utilizes 75.2% GPU memory usage. The code is available at https://github.com/LunarShen/TempMe.

Recent Blind Image Super-Resolution (BSR) methods have shown proficiency in general images. However, we find that the efficacy of recent methods obviously diminishes when employed on image data with blur, while image data with intentional blur constitute a substantial proportion of general data. To further investigate and address this issue, we developed a new super-resolution dataset specifically tailored for blur images, named the Real-world Blur-kept Super-Resolution (ReBlurSR) dataset, which consists of nearly 3000 defocus and motion blur image samples with diverse blur sizes and varying blur intensities. Furthermore, we propose a new BSR framework for blur images called Perceptual-Blur-adaptive Super-Resolution (PBaSR), which comprises two main modules: the Cross Disentanglement Module (CDM) and the Cross Fusion Module (CFM). The CDM utilizes a dual-branch parallelism to isolate conflicting blur and general data during optimization. The CFM fuses the well-optimized prior from these distinct domains cost-effectively and efficiently based on model interpolation. By integrating these two modules, PBaSR achieves commendable performance on both general and blur data without any additional inference and deployment cost and is generalizable across multiple model architectures. Rich experiments show that PBaSR achieves state-of-the-art performance across various metrics without incurring extra inference costs. Within the widely adopted LPIPS metrics, PBaSR achieves an improvement range of approximately 0.02-0.10 with diverse anchor methods and blur types, across both the ReBlurSR and multiple common general BSR benchmarks. Code here: https://github.com/Imalne/PBaSR.

Diffusion transformers (DiT) have become the de facto choice for generating high-quality images and videos, largely due to their scalability, which enables the construction of larger models for enhanced performance. However, the increased size of these models leads to higher inference costs, making them less attractive for real-time applications. We present Fast-FORward CAching (FORA), a simple yet effective approach designed to accelerate DiT by exploiting the repetitive nature of the diffusion process. FORA implements a caching mechanism that stores and reuses intermediate outputs from the attention and MLP layers across denoising steps, thereby reducing computational overhead. This approach does not require model retraining and seamlessly integrates with existing transformer-based diffusion models. Experiments show that FORA can speed up diffusion transformers several times over while only minimally affecting performance metrics such as the IS Score and FID. By enabling faster processing with minimal trade-offs in quality, FORA represents a significant advancement in deploying diffusion transformers for real-time applications. Code will be made publicly available at: https://github.com/prathebaselva/FORA.

Recent advancements in language and vision assistants have showcased impressive capabilities but suffer from a lack of transparency, limiting broader research and reproducibility. While open-source models handle general image tasks effectively, they face challenges with the high computational demands of complex visually-situated text understanding. Such tasks often require increased token inputs and large vision modules to harness high-resolution information. Striking a balance between model size and data importance remains an open question. This study aims to redefine the design of vision-language models by identifying key components and creating efficient models with constrained inference costs. By strategically formulating datasets, optimizing vision modules, and enhancing supervision techniques, we achieve significant improvements in inference throughput while maintaining high performance. Extensive experiments across models ranging from 160M to 13B parameters offer insights into model optimization. We will fully open-source our codebase, models, and datasets at https://github.com/naver-ai/elva.

In the past year, Multimodal Large Language Models (MLLMs) have demonstrated remarkable performance in tasks such as visual question answering, visual understanding and reasoning. However, the extensive model size and high training and inference costs have hindered the widespread application of MLLMs in academia and industry. Thus, studying efficient and lightweight MLLMs has enormous potential, especially in edge computing scenarios. In this survey, we provide a comprehensive and systematic review of the current state of efficient MLLMs. Specifically, we summarize the timeline of representative efficient MLLMs, research state of efficient structures and strategies, and the applications. Finally, we discuss the limitations of current efficient MLLM research and promising future directions. Please refer to our GitHub repository for more details: https://github.com/lijiannuist/Efficient-Multimodal-LLMs-Survey.

Analogical reasoning is a unique ability of humans to address unfamiliar challenges by transferring strategies from relevant past experiences. One key finding in psychology is that compared with irrelevant past experiences, recalling relevant ones can help humans better handle new tasks. Coincidentally, the NLP community has also recently found that self-generating relevant examples in the context can help large language models (LLMs) better solve a given problem than hand-crafted prompts. However, it is yet not clear whether relevance is the key factor eliciting such capability, i.e., can LLMs benefit more from self-generated relevant examples than irrelevant ones? In this work, we systematically explore whether LLMs can truly perform analogical reasoning on a diverse set of reasoning tasks. With extensive experiments and analysis, we show that self-generated random examples can surprisingly achieve comparable or even better performance, e.g., 4% performance boost on GSM8K with random biological examples. We find that the accuracy of self-generated examples is the key factor and subsequently design two improved methods with significantly reduced inference costs. Overall, we aim to advance a deeper understanding of LLM analogical reasoning and hope this work stimulates further research in the design of self-generated contexts.

Large language models (LLMs) require lengthy prompts as the input context to produce output aligned with user intentions, a process that incurs extra costs during inference. In this paper, we propose the Gist COnditioned deCOding (Gist-COCO) model, introducing a novel method for compressing prompts which also can assist the prompt interpretation and engineering. Gist-COCO employs an encoder-decoder based language model and then incorporates an additional encoder as a plugin module to compress prompts with inputs using gist tokens. It finetunes the compression plugin module and uses the representations of gist tokens to emulate the raw prompts in the vanilla language model. By verbalizing the representations of gist tokens into gist prompts, the compression ability of Gist-COCO can be generalized to different LLMs with high compression rates. Our experiments demonstrate that Gist-COCO outperforms previous prompt compression models in both passage and instruction compression tasks. Further analysis on gist verbalization results suggests that our gist prompts serve different functions in aiding language models. They may directly provide potential answers, generate the chain-of-thought, or simply repeat the inputs. All data and codes are available at https://github.com/OpenMatch/Gist-COCO .

Agents centered around Large Language Models (LLMs) are now capable of automating mobile device operations for users. After fine-tuning to learn a user's mobile operations, these agents can adhere to high-level user instructions online. They execute tasks such as goal decomposition, sequencing of sub-goals, and interactive environmental exploration, until the final objective is achieved. However, privacy concerns related to personalized user data arise during mobile operations, requiring user confirmation. Moreover, users' real-world operations are exploratory, with action data being complex and redundant, posing challenges for agent learning. To address these issues, in our practical application, we have designed interactive tasks between agents and humans to identify sensitive information and align with personalized user needs. Additionally, we integrated Standard Operating Procedure (SOP) information within the model's in-context learning to enhance the agent's comprehension of complex task execution. Our approach is evaluated on the new device control benchmark AitW, which encompasses 30K unique instructions across multi-step tasks, including application operation, web searching, and web shopping. Experimental results show that the SOP-based agent achieves state-of-the-art performance in LLMs without incurring additional inference costs, boasting an overall action success rate of 66.92\%. The code and data examples are available at https://github.com/alipay/mobile-agent.

In recent years, there have been remarkable advancements in node classification achieved by Graph Neural Networks (GNNs). However, they necessitate abundant high-quality labels to ensure promising performance. In contrast, Large Language Models (LLMs) exhibit impressive zero-shot proficiency on text-attributed graphs. Yet, they face challenges in efficiently processing structural data and suffer from high inference costs. In light of these observations, this work introduces a label-free node classification on graphs with LLMs pipeline, LLM-GNN. It amalgamates the strengths of both GNNs and LLMs while mitigating their limitations. Specifically, LLMs are leveraged to annotate a small portion of nodes and then GNNs are trained on LLMs' annotations to make predictions for the remaining large portion of nodes. The implementation of LLM-GNN faces a unique challenge: how can we actively select nodes for LLMs to annotate and consequently enhance the GNN training? How can we leverage LLMs to obtain annotations of high quality, representativeness, and diversity, thereby enhancing GNN performance with less cost? To tackle this challenge, we develop an annotation quality heuristic and leverage the confidence scores derived from LLMs to advanced node selection. Comprehensive experimental results validate the effectiveness of LLM-GNN. In particular, LLM-GNN can achieve an accuracy of 74.9% on a vast-scale dataset \products with a cost less than 1 dollar.

Person re-identification plays a key role in applications where a mobile robot needs to track its users over a long period of time, even if they are partially unobserved for some time, in order to follow them or be available on demand. In this context, deep-learning based real-time feature extraction on a mobile robot is often performed on special-purpose devices whose computational resources are shared for multiple tasks. Therefore, the inference speed has to be taken into account. In contrast, person re-identification is often improved by architectural changes that come at the cost of significantly slowing down inference. Attention blocks are one such example. We will show that some well-performing attention blocks used in the state of the art are subject to inference costs that are far too high to justify their use for mobile robotic applications. As a consequence, we propose an attention block that only slightly affects the inference speed while keeping up with much deeper networks or more complex attention blocks in terms of re-identification accuracy. We perform extensive neural architecture search to derive rules at which locations this attention block should be integrated into the architecture in order to achieve the best trade-off between speed and accuracy. Finally, we confirm that the best performing configuration on a re-identification benchmark also performs well on an indoor robotic dataset.

Transformer-based detectors (DETRs) are becoming popular for their simple framework, but the large model size and heavy time consumption hinder their deployment in the real world. While knowledge distillation (KD) can be an appealing technique to compress giant detectors into small ones for comparable detection performance and low inference cost. Since DETRs formulate object detection as a set prediction problem, existing KD methods designed for classic convolution-based detectors may not be directly applicable. In this paper, we propose DETRDistill, a novel knowledge distillation method dedicated to DETR-families. Specifically, we first design a Hungarian-matching logits distillation to encourage the student model to have the exact predictions as that of teacher DETRs. Next, we propose a target-aware feature distillation to help the student model learn from the object-centric features of the teacher model. Finally, in order to improve the convergence rate of the student DETR, we introduce a query-prior assignment distillation to speed up the student model learning from well-trained queries and stable assignment of the teacher model. Extensive experimental results on the COCO dataset validate the effectiveness of our approach. Notably, DETRDistill consistently improves various DETRs by more than 2.0 mAP, even surpassing their teacher models.

Neural painting refers to the procedure of producing a series of strokes for a given image and non-photo-realistically recreating it using neural networks. While reinforcement learning (RL) based agents can generate a stroke sequence step by step for this task, it is not easy to train a stable RL agent. On the other hand, stroke optimization methods search for a set of stroke parameters iteratively in a large search space; such low efficiency significantly limits their prevalence and practicality. Different from previous methods, in this paper, we formulate the task as a set prediction problem and propose a novel Transformer-based framework, dubbed Paint Transformer, to predict the parameters of a stroke set with a feed forward network. This way, our model can generate a set of strokes in parallel and obtain the final painting of size 512 * 512 in near real time. More importantly, since there is no dataset available for training the Paint Transformer, we devise a self-training pipeline such that it can be trained without any off-the-shelf dataset while still achieving excellent generalization capability. Experiments demonstrate that our method achieves better painting performance than previous ones with cheaper training and inference costs. Codes and models are available.

Training heuristics greatly improve various image classification model accuracies~he2018bag. Object detection models, however, have more complex neural network structures and optimization targets. The training strategies and pipelines dramatically vary among different models. In this works, we explore training tweaks that apply to various models including Faster R-CNN and YOLOv3. These tweaks do not change the model architectures, therefore, the inference costs remain the same. Our empirical results demonstrate that, however, these freebies can improve up to 5% absolute precision compared to state-of-the-art baselines.

In large vision-language models (LVLMs), images serve as inputs that carry a wealth of information. As the idiom "A picture is worth a thousand words" implies, representing a single image in current LVLMs can require hundreds or even thousands of tokens. This results in significant computational costs, which grow quadratically as input image resolution increases, thereby severely impacting the efficiency of both training and inference. Previous approaches have attempted to reduce the number of image tokens either before or within the early layers of LVLMs. However, these strategies inevitably result in the loss of crucial image information, ultimately diminishing model performance. To address this challenge, we conduct an empirical study revealing that all visual tokens are necessary for LVLMs in the shallow layers, and token redundancy progressively increases in the deeper layers of the model. To this end, we propose PyramidDrop, a visual redundancy reduction strategy for LVLMs to boost their efficiency in both training and inference with neglectable performance loss. Specifically, we partition the LVLM into several stages and drop part of the image tokens at the end of each stage with a pre-defined ratio, creating pyramid-like visual tokens across model layers. The dropping is based on a lightweight similarity calculation with a negligible time overhead. Extensive experiments demonstrate that PyramidDrop can achieve a 40% training time and 55% inference FLOPs acceleration of LLaVA-NeXT with comparable performance. Besides, the PyramidDrop could also serve as a plug-and-play strategy for inference acceleration without training, with better performance and lower inference cost than counterparts. We hope that the insights and approach introduced by PyramidDrop will inspire future research to further investigate the role of image tokens in LVLMs.

As one of the most popular parameter-efficient fine-tuning (PEFT) methods, low-rank adaptation (LoRA) is commonly applied to fine-tune large language models (LLMs). However, updating the weights of LoRA blocks effectively and expeditiously is challenging due to the long calculation path in the original model. To address this, we propose ResLoRA, an improved framework of LoRA. By adding residual paths during training and using merging approaches to eliminate these extra paths during inference, our method can achieve better results in fewer training steps without any extra trainable parameters or inference cost compared to LoRA. The experiments on NLG, NLU, and text-to-image tasks demonstrate the effectiveness of our method. To the best of our knowledge, ResLoRA is the first work that combines the residual path with LoRA. The code of our method is available at https://github.com/microsoft/LMOps/tree/main/reslora .

Large language models (LLMs) have demonstrated remarkable potential in solving complex tasks across diverse domains, typically by employing agentic workflows that follow detailed instructions and operational sequences. However, constructing these workflows requires significant human effort, limiting scalability and generalizability. Recent research has sought to automate the generation and optimization of these workflows, but existing methods still rely on initial manual setup and fall short of achieving fully automated and effective workflow generation. To address this challenge, we reformulate workflow optimization as a search problem over code-represented workflows, where LLM-invoking nodes are connected by edges. We introduce AFlow, an automated framework that efficiently explores this space using Monte Carlo Tree Search, iteratively refining workflows through code modification, tree-structured experience, and execution feedback. Empirical evaluations across six benchmark datasets demonstrate AFlow's efficacy, yielding a 5.7% average improvement over state-of-the-art baselines. Furthermore, AFlow enables smaller models to outperform GPT-4o on specific tasks at 4.55% of its inference cost in dollars. The code will be available at https://github.com/geekan/MetaGPT.

Structured state-space models (SSMs) such as S4, stemming from the seminal work of Gu et al., are gaining popularity as effective approaches for modeling sequential data. Deep SSMs demonstrate outstanding performance across a diverse set of domains, at a reduced training and inference cost compared to attention-based transformers. Recent developments show that if the linear recurrence powering SSMs allows for multiplicative interactions between inputs and hidden states (e.g. GateLoop, Mamba, GLA), then the resulting architecture can surpass in both in accuracy and efficiency attention-powered foundation models trained on text, at scales of billion parameters. In this paper, we give theoretical grounding to this recent finding using tools from Rough Path Theory: we show that when random linear recurrences are equipped with simple input-controlled transitions (selectivity mechanism), then the hidden state is provably a low-dimensional projection of a powerful mathematical object called the signature of the input -- capturing non-linear interactions between tokens at distinct timescales. Our theory not only motivates the success of modern selective state-space models such as Mamba but also provides a solid framework to understand the expressive power of future SSM variants.

In conventional object detection frameworks, a backbone body inherited from image recognition models extracts deep latent features and then a neck module fuses these latent features to capture information at different scales. As the resolution in object detection is much larger than in image recognition, the computational cost of the backbone often dominates the total inference cost. This heavy-backbone design paradigm is mostly due to the historical legacy when transferring image recognition models to object detection rather than an end-to-end optimized design for object detection. In this work, we show that such paradigm indeed leads to sub-optimal object detection models. To this end, we propose a novel heavy-neck paradigm, GiraffeDet, a giraffe-like network for efficient object detection. The GiraffeDet uses an extremely lightweight backbone and a very deep and large neck module which encourages dense information exchange among different spatial scales as well as different levels of latent semantics simultaneously. This design paradigm allows detectors to process the high-level semantic information and low-level spatial information at the same priority even in the early stage of the network, making it more effective in detection tasks. Numerical evaluations on multiple popular object detection benchmarks show that GiraffeDet consistently outperforms previous SOTA models across a wide spectrum of resource constraints. The source code is available at https://github.com/jyqi/GiraffeDet.

With the advent of Neural Radiance Fields (NeRF), neural networks can now render novel views of a 3D scene with quality that fools the human eye. Yet, generating these images is very computationally intensive, limiting their applicability in practical scenarios. In this paper, we propose a technique based on spatial decomposition capable of mitigating this issue. Our key observation is that there are diminishing returns in employing larger (deeper and/or wider) networks. Hence, we propose to spatially decompose a scene and dedicate smaller networks for each decomposed part. When working together, these networks can render the whole scene. This allows us near-constant inference time regardless of the number of decomposed parts. Moreover, we show that a Voronoi spatial decomposition is preferable for this purpose, as it is provably compatible with the Painter's Algorithm for efficient and GPU-friendly rendering. Our experiments show that for real-world scenes, our method provides up to 3x more efficient inference than NeRF (with the same rendering quality), or an improvement of up to 1.0~dB in PSNR (for the same inference cost).

As large language models (LLMs) have become increasingly compute and memory intensive, parameter-efficient fine-tuning (PEFT) methods are now a common strategy to fine-tune LLMs. A popular PEFT method is Low-Rank Adapters (LoRA), which adds trainable low-rank "adapters" to selected layers. Each adapter consists of a low-rank matrix product, multiplicatively scaled by a rank-dependent factor. This scaling factor, which divides adapters by a factor of the rank, results in slowed learning and stunted performance for LoRA with higher-rank adapters. Consequently, the use of LoRA in practice has generally been limited to very low ranks. In this work, we study the impact of the scaling factor on the learning process and prove that LoRA adapters should be divided by a factor of the square root of the rank. Modifying LoRA with the appropriate scaling factor, which we call the rank-stabilized LoRA (rsLoRA) method, easily provides for a fine-tuning compute/performance trade-off, where larger ranks can be used to trade off increased computational resources during training for better fine-tuning performance, with no change in inference computing cost.

The recent success of LLMs has been predominantly driven by curating the training dataset composition, scaling of model architectures and dataset sizes and advancements in pretraining objectives, leaving tokenizer influence as a blind spot. Shedding light on this underexplored area, we conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale, ablating different tokenizer algorithms and parameterizations. Our studies highlight that the tokenizer choice can significantly impact the model's downstream performance, training and inference costs. In particular, we find that the common tokenizer evaluation metrics fertility and parity are not always predictive of model downstream performance, rendering these metrics a questionable proxy for the model's downstream performance. Furthermore, we show that multilingual tokenizers trained on the five most frequent European languages require vocabulary size increases of factor three in comparison to English. While English-only tokenizers have been applied to the training of multi-lingual LLMs, we find that this approach results in a severe downstream performance degradation and additional training costs of up to 68%, due to an inefficient tokenization vocabulary.

Batch prompting, which combines a batch of multiple queries sharing the same context in one inference, has emerged as a promising solution to reduce inference costs. However, our study reveals a significant security vulnerability in batch prompting: malicious users can inject attack instructions into a batch, leading to unwanted interference across all queries, which can result in the inclusion of harmful content, such as phishing links, or the disruption of logical reasoning. In this paper, we construct BATCHSAFEBENCH, a comprehensive benchmark comprising 150 attack instructions of two types and 8k batch instances, to study the batch prompting vulnerability systematically. Our evaluation of both closed-source and open-weight LLMs demonstrates that all LLMs are susceptible to batch-prompting attacks. We then explore multiple defending approaches. While the prompting-based defense shows limited effectiveness for smaller LLMs, the probing-based approach achieves about 95% accuracy in detecting attacks. Additionally, we perform a mechanistic analysis to understand the attack and identify attention heads that are responsible for it.

Pre-trained Language Models (PLMs) are integral to many modern natural language processing (NLP) systems. Although multilingual models cover a wide range of languages, they often grapple with challenges like high inference costs and a lack of diverse non-English training data. Arabic-specific PLMs are trained predominantly on modern standard Arabic, which compromises their performance on regional dialects. To tackle this, we construct an Arabic dialectal corpus comprising 3.4M sentences gathered from social media platforms. We utilize this corpus to expand the vocabulary and retrain a BERT-based model from scratch. Named AlcLaM, our model was trained using only 13 GB of text, which represents a fraction of the data used by existing models such as CAMeL, MARBERT, and ArBERT, compared to 7.8%, 10.2%, and 21.3%, respectively. Remarkably, AlcLaM demonstrates superior performance on a variety of Arabic NLP tasks despite the limited training data. AlcLaM is available at GitHub https://github.com/amurtadha/Alclam and HuggingFace https://huggingface.co/rahbi.

We present Diffusion Soup, a compartmentalization method for Text-to-Image Generation that averages the weights of diffusion models trained on sharded data. By construction, our approach enables training-free continual learning and unlearning with no additional memory or inference costs, since models corresponding to data shards can be added or removed by re-averaging. We show that Diffusion Soup samples from a point in weight space that approximates the geometric mean of the distributions of constituent datasets, which offers anti-memorization guarantees and enables zero-shot style mixing. Empirically, Diffusion Soup outperforms a paragon model trained on the union of all data shards and achieves a 30% improvement in Image Reward (.34 to .44) on domain sharded data, and a 59% improvement in IR (.37 to .59) on aesthetic data. In both cases, souping also prevails in TIFA score (respectively, 85.5 to 86.5 and 85.6 to 86.8). We demonstrate robust unlearning -- removing any individual domain shard only lowers performance by 1% in IR (.45 to .44) -- and validate our theoretical insights on anti-memorization using real data. Finally, we showcase Diffusion Soup's ability to blend the distinct styles of models finetuned on different shards, resulting in the zero-shot generation of hybrid styles.

Large pre-trained language models based on transformer architecture have drastically changed the natural language processing (NLP) landscape. However, deploying those models for on-device applications in constrained devices such as smart watches is completely impractical due to their size and inference cost. As an alternative to transformer-based architectures, recent work on efficient NLP has shown that weight-efficient models can attain competitive performance for simple tasks, such as slot filling and intent classification, with model sizes in the order of the megabyte. This work introduces the pNLP-Mixer architecture, an embedding-free MLP-Mixer model for on-device NLP that achieves high weight-efficiency thanks to a novel projection layer. We evaluate a pNLP-Mixer model of only one megabyte in size on two multi-lingual semantic parsing datasets, MTOP and multiATIS. Our quantized model achieves 99.4% and 97.8% the performance of mBERT on MTOP and multi-ATIS, while using 170x fewer parameters. Our model consistently beats the state-of-the-art of tiny models (pQRNN), which is twice as large, by a margin up to 7.8% on MTOP.

Recent research on the 1-bit Large Language Models (LLMs), such as BitNet b1.58, presents a promising direction for reducing the inference cost of LLMs while maintaining their performance. In this work, we introduce BitNet a4.8, enabling 4-bit activations for 1-bit LLMs. BitNet a4.8 employs a hybrid quantization and sparsification strategy to mitigate the quantization errors introduced by the outlier channels. Specifically, we utilize 4-bit activations for inputs to the attention and feed-forward network layers, while sparsifying intermediate states followed with 8-bit quantization. Extensive experiments demonstrate that BitNet a4.8 achieves performance comparable to BitNet b1.58 with equivalent training costs, while being faster in inference with enabling 4-bit (INT4/FP4) kernels. Additionally, BitNet a4.8 activates only 55% of parameters and supports 3-bit KV cache, further enhancing the efficiency of large-scale LLM deployment and inference.

We propose a novel approach for generating complex outputs that significantly improves accuracy in text-to-SQL tasks. Our method leverages execution results to select the most semantically consistent query from multiple candidates, enabling smaller, cost-effective models to surpass computationally intensive reasoning methods such as o1, o3-mini, and DeepSeek R1 while reducing inference cost by as much as 30 times. It integrates effortlessly with existing models, offering a practical and scalable pathway to state-of-the-art SQL generation.

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3times faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. All Nemotron-H models will be released, with support in Hugging Face, NeMo, and Megatron-LM.

Although large language models (LLMs) demonstrate impressive proficiency in various tasks, they present potential safety risks, such as `jailbreaks', where malicious inputs can coerce LLMs into generating harmful content. To address these issues, many LLM developers have implemented various safety measures to align these models. This alignment involves several techniques, including data filtering during pre-training, supervised fine-tuning, reinforcement learning from human feedback, and red-teaming exercises. These methods often introduce deliberate and intentional biases similar to Political Correctness (PC) to ensure the ethical behavior of LLMs. In this paper, we delve into the intentional biases injected into LLMs for safety purposes and examine methods to circumvent these safety alignment techniques. Notably, these intentional biases result in a jailbreaking success rate in GPT-4o models that differs by 20% between non-binary and cisgender keywords and by 16% between white and black keywords, even when the other parts of the prompts are identical. We introduce the concept of PCJailbreak, highlighting the inherent risks posed by these safety-induced biases. Additionally, we propose an efficient defense method PCDefense, which prevents jailbreak attempts by injecting defense prompts prior to generation. PCDefense stands as an appealing alternative to Guard Models, such as Llama-Guard, that require additional inference cost after text generation. Our findings emphasize the urgent need for LLM developers to adopt a more responsible approach when designing and implementing safety measures.

With significant efforts in recent studies, LLM-as-a-Judge has become a cost-effective alternative to human evaluation for assessing the text generation quality in a wide range of tasks. However, there still remains a reliability gap between LLM-as-a-Judge and human evaluation. One important reason is the lack of guided oracles in the evaluation process. Motivated by the role of reference pervasively used in classic text evaluation, we introduce RevisEval, a novel text generation evaluation paradigm via the response-adapted references. RevisEval is driven by the key observation that an ideal reference should maintain the necessary relevance to the response to be evaluated. Specifically, RevisEval leverages the text revision capabilities of large language models (LLMs) to adaptively revise the response, then treat the revised text as the reference (response-adapted reference) for the subsequent evaluation. Extensive experiments demonstrate that RevisEval outperforms traditional reference-free and reference-based evaluation paradigms that use LLM-as-a-Judge across NLG tasks and open-ended instruction-following tasks. More importantly, our response-adapted references can further boost the classical text metrics, e.g., BLEU and BERTScore, compared to traditional references and even rival the LLM-as-a-Judge. A detailed analysis is also conducted to confirm RevisEval's effectiveness in bias reduction, the impact of inference cost, and reference relevance.

The reasoning-capable large language models (LLMs) demonstrate strong performance on complex reasoning tasks but often suffer from overthinking, generating unnecessarily long chain-of-thought (CoT) reasoning paths for easy reasoning questions, thereby increasing inference cost and latency. Recent approaches attempt to address this challenge by manually deciding when to apply long or short reasoning. However, they lack the flexibility to adapt CoT length dynamically based on question complexity. In this paper, we propose Auto Long-Short Reasoning (AutoL2S), a dynamic and model-agnostic framework that enables LLMs to dynamically compress their generated reasoning path based on the complexity of the reasoning question. AutoL2S enables a learned paradigm, in which LLMs themselves can decide when longer reasoning is necessary and when shorter reasoning suffices, by training on data annotated with our proposed method, which includes both long and short CoT paths and a special <EASY> token. We then use <EASY> token to indicate when the model can skip generating lengthy CoT reasoning. This proposed annotation strategy can enhance the LLMs' ability to generate shorter CoT reasoning paths with improved quality after training. Extensive evaluation results show that AutoL2S reduces the length of reasoning generation by up to 57% without compromising performance, demonstrating the effectiveness of AutoL2S for scalable and efficient LLM reasoning.

Deep learning (DL)-based code completion tools have transformed software development by enabling advanced code generation. These tools leverage models trained on vast amounts of code from numerous repositories, capturing general coding patterns. However, the impact of fine-tuning these models for specific organizations or developers to boost their performance on such subjects remains unexplored. In this work, we fill this gap by presenting solid empirical evidence answering this question. More specifically, we consider 136 developers from two organizations (Apache and Spring), two model architectures (T5 and Code Llama), and three model sizes (60M, 750M, and 7B trainable parameters). T5 models (60M, 750M) were pre-trained and fine-tuned on over 2,000 open-source projects, excluding the subject organizations' data, and compared against versions fine-tuned on organization- and developer-specific datasets. For the Code Llama model (7B), we compared the performance of the already pre-trained model publicly available online with the same model fine-tuned via parameter-efficient fine-tuning on organization- and developer-specific datasets. Our results show that there is a boost in prediction capabilities provided by both an organization-specific and a developer-specific additional fine-tuning, with the former being particularly performant. Such a finding generalizes across (i) the two subject organizations (i.e., Apache and Spring) and (ii) models of completely different magnitude (from 60M to 7B trainable parameters). Finally, we show that DL models fine-tuned on an organization-specific dataset achieve the same completion performance of pre-trained code models used out of the box and being sim10times larger, with consequent savings in terms of deployment and inference cost (e.g., smaller GPUs needed).

The training and inference of large language models (LLMs) are together a costly process that transports knowledge from raw data to meaningful computation. Inspired by the memory hierarchy of the human brain, we reduce this cost by equipping LLMs with explicit memory, a memory format cheaper than model parameters and text retrieval-augmented generation (RAG). Conceptually, with most of its knowledge externalized to explicit memories, the LLM can enjoy a smaller parameter size, training cost, and inference cost, all proportional to the amount of remaining "abstract knowledge". As a preliminary proof of concept, we train from scratch a 2.4B LLM, which achieves better performance than much larger LLMs as well as RAG models, and maintains higher decoding speed than RAG. The model is named Memory^3, since explicit memory is the third form of memory in LLMs after implicit memory (model parameters) and working memory (context key-values). We introduce a memory circuitry theory to support the externalization of knowledge, and present novel techniques including a memory sparsification mechanism that makes storage tractable and a two-stage pretraining scheme that facilitates memory formation.

Large language models (LLMs) can spend extra compute during inference to generate intermediate thoughts, which helps to produce better final responses. Since Chain-of-Thought (Wei et al., 2022), many such System 2 techniques have been proposed such as Rephrase and Respond (Deng et al., 2023a), System 2 Attention (Weston and Sukhbaatar, 2023) and Branch-Solve-Merge (Saha et al., 2023). In this work we investigate self-supervised methods to ``compile'' (distill) higher quality outputs from System 2 techniques back into LLM generations without intermediate reasoning token sequences, as this reasoning has been distilled into System 1. We show that several such techniques can be successfully distilled, resulting in improved results compared to the original System 1 performance, and with less inference cost than System 2. We posit that such System 2 distillation will be an important feature of future continually learning AI systems, enabling them to focus System 2 capabilities on the reasoning tasks that they cannot yet do well.

Knowledge distillation (KD) is widely used for compressing a teacher model to a smaller student model, reducing its inference cost and memory footprint while preserving model capabilities. However, current KD methods for auto-regressive sequence models (e.g., large language models) suffer from missing a standardized objective function. Moreover, the recent use of student-generated outputs to address training-inference mismatches has significantly escalated computational costs. To tackle these issues, we introduce DistiLLM, a more effective and efficient KD framework for auto-regressive language models. DistiLLM comprises two components: (1) a novel skew Kullback-Leibler divergence loss, where we unveil and leverage its theoretical properties, and (2) an adaptive off-policy approach designed to enhance the efficiency in utilizing student-generated outputs. Extensive experiments, including instruction-following tasks, demonstrate the effectiveness of DistiLLM in building high-performing student models while achieving up to 4.3times speedup compared to recent KD methods.

Recent research shows the potential of enhancing the problem-solving ability of large language models (LLMs) through the use of external tools. However, prior work along this line depends on the availability of existing tools. In this work, we take an initial step towards removing this dependency by proposing a closed-loop framework, referred to as LLMs As Tool Makers (LATM), where LLMs create their own reusable tools for problem-solving. Our approach consists of two key phases: 1) tool making: an LLM acts as the tool maker that crafts tools for given tasks, where a tool is implemented as a Python utility function. 2) tool using: an LLM acts as the tool user, which applies the tool built by the tool maker for problem-solving. The tool user can be either the same or a different LLM from the tool maker. Tool-making enables an LLM to continually generate tools that can be applied to different requests so that future requests can call the corresponding APIs when beneficial for solving the tasks. Furthermore, the division of labor among LLMs for tool-making and tool-using phases introduces the opportunity to achieve cost effectiveness without degrading the quality of generated tools and problem solutions. For example, recognizing that tool-making demands more sophisticated capabilities than tool-using, we can apply a powerful yet resource-intensive model as the tool maker, and a lightweight while cost-effective model as the tool user. We validate the effectiveness of our approach across a variety of complex reasoning tasks, including Big-Bench tasks. With GPT-4 as the tool maker and GPT-3.5 as the tool user, LATM can achieve performance that is on par with using GPT-4 for both tool making and tool using, while the inference cost is significantly reduced.

The growing popularity of Large Language Models has sparked interest in context compression for Large Language Models (LLMs). However, the performance of previous methods degrades dramatically as compression ratios increase, sometimes even falling to the closed-book level. This decline can be attributed to the loss of key information during the compression process. Our preliminary study supports this hypothesis, emphasizing the significance of retaining key information to maintain model performance under high compression ratios. As a result, we introduce Query-Guided Compressor (QGC), which leverages queries to guide the context compression process, effectively preserving key information within the compressed context. Additionally, we employ a dynamic compression strategy. We validate the effectiveness of our proposed QGC on the Question Answering task, including NaturalQuestions, TriviaQA, and HotpotQA datasets. Experimental results show that QGC can consistently perform well even at high compression ratios, which also offers significant benefits in terms of inference cost and throughput.

Sparsely-gated Mixture of Expert (MoE), an emerging deep model architecture, has demonstrated a great promise to enable high-accuracy and ultra-efficient model inference. Despite the growing popularity of MoE, little work investigated its potential to advance convolutional neural networks (CNNs), especially in the plane of adversarial robustness. Since the lack of robustness has become one of the main hurdles for CNNs, in this paper we ask: How to adversarially robustify a CNN-based MoE model? Can we robustly train it like an ordinary CNN model? Our pilot study shows that the conventional adversarial training (AT) mechanism (developed for vanilla CNNs) no longer remains effective to robustify an MoE-CNN. To better understand this phenomenon, we dissect the robustness of an MoE-CNN into two dimensions: Robustness of routers (i.e., gating functions to select data-specific experts) and robustness of experts (i.e., the router-guided pathways defined by the subnetworks of the backbone CNN). Our analyses show that routers and experts are hard to adapt to each other in the vanilla AT. Thus, we propose a new router-expert alternating Adversarial training framework for MoE, termed AdvMoE. The effectiveness of our proposal is justified across 4 commonly-used CNN model architectures over 4 benchmark datasets. We find that AdvMoE achieves 1% ~ 4% adversarial robustness improvement over the original dense CNN, and enjoys the efficiency merit of sparsity-gated MoE, leading to more than 50% inference cost reduction. Codes are available at https://github.com/OPTML-Group/Robust-MoE-CNN.

Chain-of-Thought (CoT) reasoning has demonstrated remarkable effectiveness in enhancing the reasoning abilities of large language models (LLMs). However, its efficiency remains a challenge due to the generation of excessive intermediate reasoning tokens, which introduce semantic redundancy and overly detailed reasoning steps. Moreover, computational expense and latency are significant concerns, as the cost scales with the number of output tokens, including those intermediate steps. In this work, we observe that most CoT tokens are unnecessary, and retaining only a small portion of them is sufficient for producing high-quality responses. Inspired by this, we propose HAWKEYE, a novel post-training and inference framework where a large model produces concise CoT instructions to guide a smaller model in response generation. HAWKEYE quantifies redundancy in CoT reasoning and distills high-density information via reinforcement learning. By leveraging these concise CoTs, HAWKEYE is able to expand responses while reducing token usage and computational cost significantly. Our evaluation shows that HAWKEYE can achieve comparable response quality using only 35% of the full CoTs, while improving clarity, coherence, and conciseness by approximately 10%. Furthermore, HAWKEYE can accelerate end-to-end reasoning by up to 3.4x on complex math tasks while reducing inference cost by up to 60%. HAWKEYE will be open-sourced and the models will be available soon.

Large language models (LLMs) like ChatGPT and GPT-4 have attracted great attention given their surprising performance on a wide range of NLP tasks. Length controlled generation of LLMs emerges as an important topic, which enables users to fully leverage the capability of LLMs in more real-world scenarios like generating a proper answer or essay of a desired length. In addition, the autoregressive generation in LLMs is extremely time-consuming, while the ability of controlling this generated length can reduce the inference cost by limiting the length. Therefore, we propose a prompt-based length control method to achieve high-accuracy length controlled generation. In particular, we adopt reinforcement learning with the reward signal given by either trainable or rule-based reward models, which further enhances the length-control ability of LLMs by rewarding outputs that follows pre-defined control instruction. To enable rule-based inference, we also introduce standard prompt extractor to collect the standard control information from users' input. Experiments show that our method significantly improves the accuracy of prompt-based length control for summarization task on popular datasets like CNNDM and NYT. Both the standard prompt extractor and the RL-tuned model have show strong generalization ability to unseen control prompt templates.

The Abstraction and Reasoning Corpus (ARC-AGI) poses a significant challenge for large language models (LLMs), exposing limitations in their abstract reasoning abilities. In this work, we leverage task-specific data augmentations throughout the training, generation, and scoring phases, and employ a depth-first search algorithm to generate diverse, high-probability candidate solutions. Furthermore, we utilize the LLM not only as a generator but also as a scorer, using its output probabilities to select the most promising solutions. Our method achieves a score of 71.6% (286.5/400 solved tasks) on the public ARC-AGI evaluation set, demonstrating state-of-the-art performance among publicly available approaches. While concurrent closed-source work has reported higher scores, our method distinguishes itself through its transparency, reproducibility, and remarkably low inference cost, averaging only around 2ct per task on readily available hardware (we assume a price of 36ct/hour for a Nvidia 4090 GPU).

Aligning diffusion models to downstream tasks often requires finetuning new models or gradient-based guidance at inference time to enable sampling from the reward-tilted posterior. In this work, we explore a simple inference-time gradient-free guidance approach, called controlled denoising (CoDe), that circumvents the need for differentiable guidance functions and model finetuning. CoDe is a blockwise sampling method applied during intermediate denoising steps, allowing for alignment with downstream rewards. Our experiments demonstrate that, despite its simplicity, CoDe offers a favorable trade-off between reward alignment, prompt instruction following, and inference cost, achieving a competitive performance against the state-of-the-art baselines. Our code is available at: https://github.com/anujinho/code.

In search scenarios, user experience can be hindered by erroneous queries due to typos, voice errors, or knowledge gaps. Therefore, query correction is crucial for search engines. Current correction models, usually small models trained on specific data, often struggle with queries beyond their training scope or those requiring contextual understanding. While the advent of Large Language Models (LLMs) offers a potential solution, they are still limited by their pre-training data and inference cost, particularly for complex queries, making them not always effective for query correction. To tackle these, we propose Trigger^3, a large-small model collaboration framework that integrates the traditional correction model and LLM for query correction, capable of adaptively choosing the appropriate correction method based on the query and the correction results from the traditional correction model and LLM. Trigger^3 first employs a correction trigger to filter out correct queries. Incorrect queries are then corrected by the traditional correction model. If this fails, an LLM trigger is activated to call the LLM for correction. Finally, for queries that no model can correct, a fallback trigger decides to return the original query. Extensive experiments demonstrate Trigger^3 outperforms correction baselines while maintaining efficiency.

While significant progress has been made on the text-to-SQL task, recent solutions repeatedly encode the same database schema for every question, resulting in unnecessary high inference cost and often overlooking crucial database knowledge. To address these issues, we propose You Only Read Once (YORO), a novel paradigm that directly internalizes database knowledge into the parametric knowledge of a text-to-SQL model during training and eliminates the need for schema encoding during inference. YORO significantly reduces the input token length by 66%-98%. Despite its shorter inputs, our empirical results demonstrate YORO's competitive performances with traditional systems on three benchmarks as well as its significant outperformance on large databases. Furthermore, YORO excels in handling questions with challenging value retrievals such as abbreviation.

Managing long texts is challenging for large language models (LLMs) due to limited context window sizes. This study introduces UIO-LLMs, an unbiased incremental optimization approach for memory-enhanced transformers under long-context settings. We initially conceptualize the process as a streamlined encoder-decoder framework where the weights-shared encoder and decoder respectively encapsulate a context segment into memories and leverage these memories to predict outputs of the subsequent segment. Subsequently, by treating our memory-enhanced transformers as fully-connected recurrent neural networks (RNNs), we refine the training process using the Truncated Backpropagation Through Time (TBPTT) algorithm, which incorporates innovative incremental optimization techniques. These techniques not only diminish time complexity but also address the bias in gradient computation through an unbiased optimization process. UIO-LLMs successfully handle long context, such as extending the context window of Llama2-7b-chat from 4K to 100K tokens with minimal 2% additional parameters, while keeping the inference cost nearly linear as context length increases.

Transformer-based long context generative models power emerging AI applications like hour-long video understanding and project-level coding agent. Deploying long context transformers (e.g., 100K to 10M tokens) is prohibitively expensive compared to short context (e.g., 4K tokens) model variants. Reducing the cost of long-context transformers is becoming a pressing research and engineering challenge starting from the year of 2024. This work describes a concurrent programming framework for quantitatively analyzing the efficiency challenges in serving multiple long-context requests under limited size of GPU high-bandwidth memory (HBM) regime. We give a detailed analysis of how all additional computational costs, compared to 4K context, trace back to one single source: the large size of the KV cache. We use a 34B GPT-3.5 level model of 50K context on A100 NVLink as a running example, and describe how its large KV cache causes four types of deployment challenges: (1) prefilling long inputs takes much longer compute time and GPU memory than short inputs; (2) after prefilling, the large KV cache residing on the GPU HBM substantially restricts the number of concurrent users being served; (3) during decoding, repeatedly reading the KV cache from HBM to SM largely increases latency; (4) when KV cache memory overflows, swapping it from HBM to DDR causes significant context switching latency. We use this framework to analyze existing works and identify possibilities of combining them to build end-to-end systems. Overall, this work offers a foundational framework for analyzing long context transformer deployment and identifies directions towards reducing the inference cost of 1M context to be as cheap as 4K.

Despite the recent advancements, conditional image generation still faces challenges of cost, generalizability, and the need for task-specific training. In this paper, we propose Manifold Preserving Guided Diffusion (MPGD), a training-free conditional generation framework that leverages pretrained diffusion models and off-the-shelf neural networks with minimal additional inference cost for a broad range of tasks. Specifically, we leverage the manifold hypothesis to refine the guided diffusion steps and introduce a shortcut algorithm in the process. We then propose two methods for on-manifold training-free guidance using pre-trained autoencoders and demonstrate that our shortcut inherently preserves the manifolds when applied to latent diffusion models. Our experiments show that MPGD is efficient and effective for solving a variety of conditional generation applications in low-compute settings, and can consistently offer up to 3.8x speed-ups with the same number of diffusion steps while maintaining high sample quality compared to the baselines.

We introduce Tangent Attention Fine-Tuning (TAFT), a method for fine-tuning linearized transformers obtained by computing a First-order Taylor Expansion around a pre-trained initialization. We show that the Jacobian-Vector Product resulting from linearization can be computed efficiently in a single forward pass, reducing training and inference cost to the same order of magnitude as its original non-linear counterpart, while using the same number of parameters. Furthermore, we show that, when applied to various downstream visual classification tasks, the resulting Tangent Transformer fine-tuned with TAFT can perform comparably with fine-tuning the original non-linear network. Since Tangent Transformers are linear with respect to the new set of weights, and the resulting fine-tuning loss is convex, we show that TAFT enjoys several advantages compared to non-linear fine-tuning when it comes to model composition, parallel training, machine unlearning, and differential privacy.

Convolutional Neural Networks (CNNs) are the go-to model for computer vision. Recently, attention-based networks, such as the Vision Transformer, have also become popular. In this paper we show that while convolutions and attention are both sufficient for good performance, neither of them are necessary. We present MLP-Mixer, an architecture based exclusively on multi-layer perceptrons (MLPs). MLP-Mixer contains two types of layers: one with MLPs applied independently to image patches (i.e. "mixing" the per-location features), and one with MLPs applied across patches (i.e. "mixing" spatial information). When trained on large datasets, or with modern regularization schemes, MLP-Mixer attains competitive scores on image classification benchmarks, with pre-training and inference cost comparable to state-of-the-art models. We hope that these results spark further research beyond the realms of well established CNNs and Transformers.

Vision-language models (VLMs) pretrained on large-scale multimodal datasets encode rich visual and linguistic knowledge, making them a strong foundation for robotics. Rather than training robotic policies from scratch, recent approaches adapt VLMs into vision-language-action (VLA) models that enable natural language-driven perception and control. However, existing VLAs are typically massive--often with billions of parameters--leading to high training costs and limited real-world deployability. Moreover, they rely on academic and industrial datasets, overlooking the growing availability of community-collected data from affordable robotic platforms. In this work, we present SmolVLA, a small, efficient, and community-driven VLA that drastically reduces both training and inference costs, while retaining competitive performance. SmolVLA is designed to be trained on a single GPU and deployed on consumer-grade GPUs or even CPUs. To further improve responsiveness, we introduce an asynchronous inference stack decoupling perception and action prediction from action execution, allowing higher control rates with chunked action generation. Despite its compact size, SmolVLA achieves performance comparable to VLAs that are 10x larger. We evaluate SmolVLA on a range of both simulated as well as real-world robotic benchmarks and release all code, pretrained models, and training data.

This technical report presents Yi-Lightning, our latest flagship large language model (LLM). It achieves exceptional performance, ranking 6th overall on Chatbot Arena, with particularly strong results (2nd to 4th place) in specialized categories including Chinese, Math, Coding, and Hard Prompts. Yi-Lightning leverages an enhanced Mixture-of-Experts (MoE) architecture, featuring advanced expert segmentation and routing mechanisms coupled with optimized KV-caching techniques. Our development process encompasses comprehensive pre-training, supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF), where we devise deliberate strategies for multi-stage training, synthetic data construction, and reward modeling. Furthermore, we implement RAISE (Responsible AI Safety Engine), a four-component framework to address safety issues across pre-training, post-training, and serving phases. Empowered by our scalable super-computing infrastructure, all these innovations substantially reduce training, deployment and inference costs while maintaining high-performance standards. With further evaluations on public academic benchmarks, Yi-Lightning demonstrates competitive performance against top-tier LLMs, while we observe a notable disparity between traditional, static benchmark results and real-world, dynamic human preferences. This observation prompts a critical reassessment of conventional benchmarks' utility in guiding the development of more intelligent and powerful AI systems for practical applications. Yi-Lightning is now available through our developer platform at https://platform.lingyiwanwu.com.

Among the widely used parameter-efficient finetuning (PEFT) methods, LoRA and its variants have gained considerable popularity because of avoiding additional inference costs. However, there still often exists an accuracy gap between these methods and full fine-tuning (FT). In this work, we first introduce a novel weight decomposition analysis to investigate the inherent differences between FT and LoRA. Aiming to resemble the learning capacity of FT from the findings, we propose Weight-Decomposed LowRank Adaptation (DoRA). DoRA decomposes the pre-trained weight into two components, magnitude and direction, for fine-tuning, specifically employing LoRA for directional updates to efficiently minimize the number of trainable parameters. By employing DoRA, we enhance both the learning capacity and training stability of LoRA while avoiding any additional inference overhead. DoRA consistently outperforms LoRA on fine-tuning LLaMA, LLaVA, and VL-BART on various downstream tasks, such as commonsense reasoning, visual instruction tuning, and image/video-text understanding.

Large Reasoning Models (LRMs) excel at complex tasks using Chain-of-Thought (CoT) reasoning. However, their tendency to overthinking leads to unnecessarily lengthy reasoning chains, dramatically increasing inference costs. To mitigate this issue, we introduce VeriThinker, a novel approach for CoT compression. Unlike conventional methods that fine-tune LRMs directly on the original reasoning task using synthetic concise CoT data, we innovatively fine-tune the model solely through an auxiliary verification task. By training LRMs to accurately verify the correctness of CoT solutions, the LRMs inherently become more discerning about the necessity of subsequent self-reflection steps, thereby effectively suppressing overthinking. Extensive experiments validate that VeriThinker substantially reduces reasoning chain lengths while maintaining or even slightly improving accuracy. When applied to DeepSeek-R1-Distill-Qwen-7B, our approach reduces reasoning tokens on MATH500 from 3790 to 2125 while improving accuracy by 0.8% (94.0% to 94.8%), and on AIME25, tokens decrease from 14321 to 10287 with a 2.1% accuracy gain (38.7% to 40.8%). Additionally, our experiments demonstrate that VeriThinker can also be zero-shot generalized to speculative reasoning. Code is available at https://github.com/czg1225/VeriThinker

We propose SCONE (Scalable, Contextualized, Offloaded, N-gram Embedding), a method for extending input embedding layers to enhance language model performance as layer size scales. To avoid increased decoding costs, SCONE retains the original vocabulary while introducing embeddings for a set of frequent n-grams. These embeddings provide contextualized representation for each input token and are learned with a separate model during training. During inference, they are precomputed and stored in off-accelerator memory with minimal impact on inference speed. SCONE enables two new scaling strategies: increasing the number of cached n-gram embeddings and scaling the model used to learn them, all while maintaining fixed inference-time FLOPS. We show that scaling both aspects allows SCONE to outperform a 1.9B parameter baseline across diverse corpora, while using only half the inference-time FLOPS.

Evaluating large language models (LLMs) effectively remains a critical bottleneck, as traditional static benchmarks suffer from saturation and contamination, while human evaluations are costly and slow. This hinders timely or domain-specific assessment, crucial for real-world applications. We introduce YourBench, a novel, open-source framework that addresses these limitations by enabling dynamic, automated generation of reliable, up-to-date, and domain-tailored benchmarks cheaply and without manual annotation, directly from user-provided documents. We demonstrate its efficacy by replicating 7 diverse MMLU subsets using minimal source text, achieving this for under 15 USD in total inference costs while perfectly preserving the relative model performance rankings (Spearman Rho = 1) observed on the original benchmark. To ensure that YourBench generates data grounded in provided input instead of relying on posterior parametric knowledge in models, we also introduce Tempora-0325, a novel dataset of over 7K diverse documents, published exclusively after March 2025. Our comprehensive analysis spans 26 SoTA models from 7 major families across varying scales (3-671B parameters) to validate the quality of generated evaluations through rigorous algorithmic checks (e.g., citation grounding) and human assessments. We release the YourBench library, the Tempora-0325 dataset, 150k+ question answer pairs based on Tempora and all evaluation and inference traces to facilitate reproducible research and empower the community to generate bespoke benchmarks on demand, fostering more relevant and trustworthy LLM evaluation.

We propose Diffusion-Sharpening, a fine-tuning approach that enhances downstream alignment by optimizing sampling trajectories. Existing RL-based fine-tuning methods focus on single training timesteps and neglect trajectory-level alignment, while recent sampling trajectory optimization methods incur significant inference NFE costs. Diffusion-Sharpening overcomes this by using a path integral framework to select optimal trajectories during training, leveraging reward feedback, and amortizing inference costs. Our method demonstrates superior training efficiency with faster convergence, and best inference efficiency without requiring additional NFEs. Extensive experiments show that Diffusion-Sharpening outperforms RL-based fine-tuning methods (e.g., Diffusion-DPO) and sampling trajectory optimization methods (e.g., Inference Scaling) across diverse metrics including text alignment, compositional capabilities, and human preferences, offering a scalable and efficient solution for future diffusion model fine-tuning. Code: https://github.com/Gen-Verse/Diffusion-Sharpening

Recent approaches have shown promises distilling diffusion models into efficient one-step generators. Among them, Distribution Matching Distillation (DMD) produces one-step generators that match their teacher in distribution, without enforcing a one-to-one correspondence with the sampling trajectories of their teachers. However, to ensure stable training, DMD requires an additional regression loss computed using a large set of noise-image pairs generated by the teacher with many steps of a deterministic sampler. This is costly for large-scale text-to-image synthesis and limits the student's quality, tying it too closely to the teacher's original sampling paths. We introduce DMD2, a set of techniques that lift this limitation and improve DMD training. First, we eliminate the regression loss and the need for expensive dataset construction. We show that the resulting instability is due to the fake critic not estimating the distribution of generated samples accurately and propose a two time-scale update rule as a remedy. Second, we integrate a GAN loss into the distillation procedure, discriminating between generated samples and real images. This lets us train the student model on real data, mitigating the imperfect real score estimation from the teacher model, and enhancing quality. Lastly, we modify the training procedure to enable multi-step sampling. We identify and address the training-inference input mismatch problem in this setting, by simulating inference-time generator samples during training time. Taken together, our improvements set new benchmarks in one-step image generation, with FID scores of 1.28 on ImageNet-64x64 and 8.35 on zero-shot COCO 2014, surpassing the original teacher despite a 500X reduction in inference cost. Further, we show our approach can generate megapixel images by distilling SDXL, demonstrating exceptional visual quality among few-step methods.

Scaling laws are useful guides for developing language models, but there are still gaps between current scaling studies and how language models are ultimately trained and evaluated. For instance, scaling is usually studied in the compute-optimal training regime (i.e., "Chinchilla optimal" regime); however, in practice, models are often over-trained to reduce inference costs. Moreover, scaling laws mostly predict loss on next-token prediction, but ultimately models are compared based on downstream task performance. In this paper, we address both shortcomings. To do so, we create a testbed of 104 models with 0.011B to 6.9B parameters trained with various numbers of tokens on three data distributions. First, we investigate scaling in the over-trained regime. We fit scaling laws that extrapolate in both the number of model parameters and the ratio of training tokens to parameters. This enables us to predict the validation loss of a 1.4B parameter, 900B token run (i.e., 32times over-trained) and a 6.9B parameter, 138B token runx2014each from experiments that take 300times less compute. Second, we relate the perplexity of a language model to its downstream task performance via a power law. We use this law to predict top-1 error averaged over downstream tasks for the two aforementioned models using experiments that take 20times less compute. Our experiments are available at https://github.com/mlfoundations/scaling.

Chain-of-Thought significantly enhances a model's reasoning capability, but it also comes with a considerable increase in inference costs due to long chains. With the observation that the reasoning path can be easily compressed under easy tasks but struggle on hard tasks, we explore the feasibility of elastically controlling the length of reasoning paths with only one model, thereby reducing the inference overhead of reasoning models dynamically based on task difficulty. We introduce a new tuning and inference strategy named CoT-Valve, designed to allow models to generate reasoning chains of varying lengths. To achieve this, we propose to identify a direction in the parameter space that, when manipulated, can effectively control the length of generated CoT. Moreover, we show that this property is valuable for compressing the reasoning chain. We construct datasets with chains from long to short for the same questions and explore two enhanced strategies for CoT-Valve: (1) a precise length-compressible CoT tuning method, and (2) a progressive chain length compression approach. Our experiments show that CoT-Valve successfully enables controllability and compressibility of the chain and shows better performance than the prompt-based control. We applied this method to QwQ-32B-Preview, reducing reasoning chains on GSM8K from 741 to 225 tokens with a minor performance drop (95.07% to 94.92%) and on AIME from 6827 to 4629 tokens, with only one additional incorrect answer.

We propose to improve transformers of a specific modality with irrelevant data from other modalities, e.g., improve an ImageNet model with audio or point cloud datasets. We would like to highlight that the data samples of the target modality are irrelevant to the other modalities, which distinguishes our method from other works utilizing paired (e.g., CLIP) or interleaved data of different modalities. We propose a methodology named Multimodal Pathway - given a target modality and a transformer designed for it, we use an auxiliary transformer trained with data of another modality and construct pathways to connect components of the two models so that data of the target modality can be processed by both models. In this way, we utilize the universal sequence-to-sequence modeling abilities of transformers obtained from two modalities. As a concrete implementation, we use a modality-specific tokenizer and task-specific head as usual but utilize the transformer blocks of the auxiliary model via a proposed method named Cross-Modal Re-parameterization, which exploits the auxiliary weights without any inference costs. On the image, point cloud, video, and audio recognition tasks, we observe significant and consistent performance improvements with irrelevant data from other modalities. The code and models are available at https://github.com/AILab-CVC/M2PT.

While the Mamba architecture demonstrates superior inference efficiency and competitive performance on short-context natural language processing (NLP) tasks, empirical evidence suggests its capacity to comprehend long contexts is limited compared to transformer-based models. In this study, we investigate the long-context efficiency issues of the Mamba models and propose ReMamba, which enhances Mamba's ability to comprehend long contexts. ReMamba incorporates selective compression and adaptation techniques within a two-stage re-forward process, incurring minimal additional inference costs overhead. Experimental results on the LongBench and L-Eval benchmarks demonstrate ReMamba's efficacy, improving over the baselines by 3.2 and 1.6 points, respectively, and attaining performance almost on par with same-size transformer models.

Diffusion Policies have become widely used in Imitation Learning, offering several appealing properties, such as generating multimodal and discontinuous behavior. As models are becoming larger to capture more complex capabilities, their computational demands increase, as shown by recent scaling laws. Therefore, continuing with the current architectures will present a computational roadblock. To address this gap, we propose Mixture-of-Denoising Experts (MoDE) as a novel policy for Imitation Learning. MoDE surpasses current state-of-the-art Transformer-based Diffusion Policies while enabling parameter-efficient scaling through sparse experts and noise-conditioned routing, reducing both active parameters by 40% and inference costs by 90% via expert caching. Our architecture combines this efficient scaling with noise-conditioned self-attention mechanism, enabling more effective denoising across different noise levels. MoDE achieves state-of-the-art performance on 134 tasks in four established imitation learning benchmarks (CALVIN and LIBERO). Notably, by pretraining MoDE on diverse robotics data, we achieve 4.01 on CALVIN ABC and 0.95 on LIBERO-90. It surpasses both CNN-based and Transformer Diffusion Policies by an average of 57% across 4 benchmarks, while using 90% fewer FLOPs and fewer active parameters compared to default Diffusion Transformer architectures. Furthermore, we conduct comprehensive ablations on MoDE's components, providing insights for designing efficient and scalable Transformer architectures for Diffusion Policies. Code and demonstrations are available at https://mbreuss.github.io/MoDE_Diffusion_Policy/.

Large Language Models (LLMs) such as ChatGPT and LlaMA are advancing rapidly in generative Artificial Intelligence (AI), but their immense size poses significant challenges, such as huge training and inference costs, substantial energy demands, and limitations for on-site deployment. Traditional compression methods such as pruning, distillation, and low-rank approximation focus on reducing the effective number of neurons in the network, while quantization focuses on reducing the numerical precision of individual weights to reduce the model size while keeping the number of neurons fixed. While these compression methods have been relatively successful in practice, there is no compelling reason to believe that truncating the number of neurons is an optimal strategy. In this context, this paper introduces CompactifAI, an innovative LLM compression approach using quantum-inspired Tensor Networks that focuses on the model's correlation space instead, allowing for a more controlled, refined and interpretable model compression. Our method is versatile and can be implemented with - or on top of - other compression techniques. As a benchmark, we demonstrate that a combination of CompactifAI with quantization allows to reduce a 93% the memory size of LlaMA 7B, reducing also 70% the number of parameters, accelerating 50% the training and 25% the inference times of the model, and just with a small accuracy drop of 2% - 3%, going much beyond of what is achievable today by other compression techniques. Our methods also allow to perform a refined layer sensitivity profiling, showing that deeper layers tend to be more suitable for tensor network compression, which is compatible with recent observations on the ineffectiveness of those layers for LLM performance. Our results imply that standard LLMs are, in fact, heavily overparametrized, and do not need to be large at all.

Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. The majority of computation stems from the overwhelming volume of vision tokens processed by the transformer decoder. In this paper, we propose to build efficient MLLMs by leveraging the Mixture-of-Depths (MoD) mechanism, where each transformer decoder layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layer and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. To validate the effectiveness of our approach, we conduct extensive experiments with two baseline models across 14 benchmarks. Our model, p-MoD, matches or even surpasses the performance of the baseline models, with only 55.6% TFLOPs and 53.8% KV cache storage during inference, and 77.7% GPU hours during training.

Vision-language foundation models (such as CLIP) have recently shown their power in transfer learning, owing to large-scale image-text pre-training. However, target domain data in the downstream tasks can be highly different from the pre-training phase, which makes it hard for such a single model to generalize well. Alternatively, there exists a wide range of expert models that contain diversified vision and/or language knowledge pre-trained on different modalities, tasks, networks, and datasets. Unfortunately, these models are "isolated agents" with heterogeneous structures, and how to integrate their knowledge for generalizing CLIP-like models has not been fully explored. To bridge this gap, we propose a general and concise TransAgent framework, which transports the knowledge of the isolated agents in a unified manner, and effectively guides CLIP to generalize with multi-source knowledge distillation. With such a distinct framework, we flexibly collaborate with 11 heterogeneous agents to empower vision-language foundation models, without further cost in the inference phase. Finally, our TransAgent achieves state-of-the-art performance on 11 visual recognition datasets. Under the same low-shot setting, it outperforms the popular CoOp with around 10% on average, and 20% on EuroSAT which contains large domain shifts.

Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.