SmartThinker: Progressive Chain-of-Thought Length Calibration for Efficient Large Language Model Reasoning

TL;DR

SmartThinker introduces a dynamic chain-of-thought length calibration method for large reasoning models, reducing verbosity and improving accuracy by adaptively optimizing response length during training.

Contribution

It proposes a novel GRPO-based approach that dynamically estimates optimal reasoning length and adjusts reward coefficients to enhance efficiency and accuracy.

Findings

Achieves up to 52.5% length compression with better accuracy.

Improves accuracy by up to 16.6% on challenging benchmarks.

Effectively balances response length and reasoning quality.

Abstract

Large reasoning models (LRMs) like OpenAI o1 and DeepSeek-R1 achieve high accuracy on complex tasks by adopting long chain-of-thought (CoT) reasoning paths. However, the inherent verbosity of these processes frequently results in redundancy and overthinking. To address this issue, existing works leverage Group Relative Policy Optimization (GRPO) to reduce LRM output length, but their static length reward design cannot dynamically adapt according to the relative problem difficulty and response length distribution, causing over-compression and compromised accuracy. Therefore, we propose SmartThinker, a novel GRPO-based efficient reasoning method with progressive CoT length calibration. SmartThinker makes a two-fold contribution: First, it dynamically estimates the optimal length with peak accuracy during training and guides overlong responses toward it to reduce response length while sustaining accuracy. Second, it dynamically modulates the length reward coefficient to avoid the unwarranted penalization of correct reasoning paths. Extensive experiment results show that SmartThinker achieves up to 52.5% average length compression with improved accuracy, and achieves up to 16.6% accuracy improvement on challenging benchmarks like AIME25. The source code can be found at https://github.com/SJTU-RTEAS/SmartThinker.