THOUGHTTERMINATOR: Benchmarking, Calibrating, and Mitigating Overthinking in Reasoning Models
saxon
Abstract
Reasoning models have demonstrated impressive performance on difficult tasks that traditional language models struggle at. However, many are plagued with the problem of overthinking--generating large amounts of unnecessary tokens which don't improve accuracy on a question. We introduce approximate measures of problem-level difficulty and demonstrate that a clear relationship between problem difficulty and optimal token spend exists, and evaluate how well calibrated a variety of reasoning models are in terms of efficiently allocating the optimal token count. We find that in general, reasoning models are poorly calibrated, particularly on easy problems. To evaluate calibration on easy questions we introduce DUMB500, a dataset of extremely easy math, reasoning, code, and task problems, and jointly evaluate reasoning model on these simple examples and extremely difficult examples from existing frontier benchmarks on the same task domain. Finally, we introduce THOUGHTTERMINATOR, a training-free black box decoding technique that significantly improves reasoning model calibration.
Community
From a simple observational measure of overthinking, we introduce Thought Terminator, a black-box, training-free decoding technique where RMs set their own deadlines and follow them.
Our measure of overthinking is stupid simple: what's the delta between the mean/max token spend on each question vs the minimum for successful answers.
There exists a clear trend between question difficulty (measured by success rates) and required spend. In order to sample a more balanced distribution of questions across the difficulty spectrum, we introduce DUMB500, the Waluigi to MATH500 which consists of stupid easy Qs.
This way we can get a more comprehensive view of overthinking, from the hardest GPQA and ZebraLogic Qs to literally "2+2=?"
Finally, we introduce Thought Terminator, our Schwarzeneggerian method to mitigating overthinking, which is a modified decoder that inserts interrupts every N tokens to tell the model how much compute it has left. Once that budget is spent it uses constrained decoding for budget forcing. Across multiple RMs, Terminator calibrates performance, getting near-optimal performance in significantly fewer tokens.
Most interestingly, our model-predicted deadlines find the OPTIMAL budget, near the plateau where further spend isn't beneficial
In this way Terminator is a tool any RM can use!
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Self-Training Elicits Concise Reasoning in Large Language Models (2025)
- DNR Bench: Benchmarking Over-Reasoning in Reasoning LLMs (2025)
- Leveraging Reasoning Model Answers to Enhance Non-Reasoning Model Capability (2025)
- Pensez: Less Data, Better Reasoning - Rethinking French LLM (2025)
- DeduCE: Deductive Consistency as a Framework to Evaluate LLM Reasoning (2025)
- m1: Unleash the Potential of Test-Time Scaling for Medical Reasoning with Large Language Models (2025)
- Speculative Thinking: Enhancing Small-Model Reasoning with Large Model Guidance at Inference Time (2025)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper