Chain-of-thought reasoning improves accuracy but dramatically increases token consumption. This paper finds that tokens within reasoning traces are not equally valuable: core tokens carry decisive content, while redundant tokens are exploratory filler. They propose TokenProbe, which uses normalized log probability to distinguish these tokens, and apply a GRPO training objective to compress redundant ones. This reduces token usage by 76% compared to baseline without harming reasoning quality. Under matched reasoning-length budgets, their method even outperforms Gemini-3.1-Pro, a strong closed-source baseline. The work is interesting because it directly addresses the efficiency-accuracy trade-off in reasoning, potentially enabling cheaper and faster reasoning in LLMs.
Language models can cut reasoning tokens by targeting low-value text
TokenProbe uses token-level confidence signals during reinforcement learning to compress reasoning traces while preserving reported task accuracy.
Big Tech
Runjia Zeng · Hang Hua · Yiyang Liu · Zhiqiang Tao · Ruixiang Tang · Qifan Wang · +2 more
Purdue University · Rochester Institute of Technology · MIT-IBM Watson AI Lab · University of Missouri-Kansas City · Rutgers University
Research Digest··2 min read
Thread:Token Pruning Efficiency
In one line
Normalized log probability identifies core and redundant tokens; compressing redundant tokens reduces token usage by 76% while preserving reasoning quality.
Before you read this
An idea this paper builds onRead it as
Zeng and colleagues examine whether every token in a chain-of-thought trace contributes equally to solving a problem.
The authors analyzed reasoning traces at token level, using normalized log probability, a model-relative measure of how confidently each token was generated.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.