Note General model is not great at specializing tasks. Narrow-domain fine-tuned checkpoint becomes better at specific tasks, such local improvement can feedback onto the full training dataset, achieving self-augmentation based improvement. This is a interesting idea.

Note Use small language model to search the graph and route to the doman expert.

Note Automatic Flow Engineering done by 3B fine-tuned LLM, grounded on selective set of API-based functions. Planning model perform task decomposition, but do not do specific calls. Effectively doing flow (prompt) engineering here. Topology in plans are lacking and static plan-ahead approach is less robust (although good according to their curated 1k test dataset)

Note Treating intermediate embedding sequences as a bunch of signals and apply 1D convolution on temporal axis, similar to ConvMixer's manipulation in some sense, experimentation conducted on pre-training transformer. Interesting result is reported in the paper. Unfortunately no 'wave' is actually applied, no 'periodic' information is captured.

Note Newton's introduction of gravity illustrates how understanding derivatives—knowing how things move rather than just where they are—enhances reasoning about the world. Large language models (LLMs), while excelling at compressing data distributions, struggle with reasoning. Reasoning involves grasping the 'abstract structure' of data. Therefore, by modeling derivatives of data distributions, could we improve LLMs' reasoning capabilities?

Note Tokenization unifies perception and generation, end-to-end training with discrete multi-modality signal enables both.

Note Intelligence is very likely the ability to model higher order derivatives given lower order observation.

Note MLLM usually project a continuous Image embedding onto hidden space of LLM. Vector quantization (VQ) convert an image into discrete codes representing each of its patches, these tokens could be ported into LLM in a more similar fashion as text tokens -- new embedding vectors. Therefore a natural extension is just to re-use the BPE approach onto these image tokens. Which is precisely what happens in this work. However, I

Note "If two computer programs perform the same task, the shorter one is generally better." This principle, known as Occam's Razor, is a critical guideline for scientific discovery. Our best program today is the Transformer. Can we make it more efficient? Selective attention improves the Transformer by allowing each token to decide whether previous context is still relevant for future tokens.

Note I find it strange to view encoder which produces embedding vector as a type of tokenization --- then transformer effectively has two tokenization process... a discrete one and then a continuous one ?

Note Compression requires redundancy, otherwise it's just memorization

Note Encodes interaction trajectories into "skill vectors" that act like abstract concepts: a skill decoder (low-level policy) translates them into specific actions based on the current state—similar to how our concepts become concrete actions in different situations. By relabeling experiences with these skills, they train a high-level policy to select optimal skills that maximize rewards. This hierarchical approach hints at the possibility for AI systems to formulate and think in their own-curated a

Note Random network distillation as extra reward for exploration encouragement for RL.

Note Re-using previous experience to increase RL learning efficiency.

Note Using fixed tokens to encode image, adding new tokens recursively until reaching satisfacotry compression level.