Reproduce “ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy” on OpenResearch

About this paper

Gengyang Li, Yifeng Gao, Yuming Li, Yunfang Wu

Peking University researchers developed ThinkLess, a training-free and inference-efficient method that reduces reasoning redundancy in large language models. This approach maintains or enhances accuracy compared to full Chain-of-Thought, while significantly decreasing token usage by 60-70% and inference time by over 50%.

Key insights

• LLMs exhibit a 'progressive information migration' where reasoning content is continuously summarized and compressed into later tokens, with earlier steps becoming less critical for the final answer.
• The `</think>` token acts as a 'semantic anchor,' implicitly encoding a compressed, internalized representation of the reasoning process, suggesting that explicit CoT steps are often a verbose articulation of an already-formed internal state.
• Naive early termination preserves internal reasoning but requires an output regulation mechanism to ensure correct answer formatting and maintain high evaluated accuracy.

Summary by alphaXiv.