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

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 designs fail to adapt to problem difficulty and response-length distributions, 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 reasoning length while sustaining accuracy. Second, it dynamically modulates the length-reward coefficient to avoid the unwarranted penalization of correct reasoning paths. Extensive experimental results show that SmartThinker achieves up to 52.6\% length compression with improved accuracy and achieves up to 16.6\% accuracy relative improvement on challenging benchmarks like AIME25. The source code can be found at https://github.com/SJTU-RTEAS/SmartThinker.

This paper studies how to make large language models answer complex questions more efficiently and accurately. Current models often improve their answers by producing long step-by-step reasoning, but much of this reasoning can be repetitive, unnecessary, or overly complicated. This increases waiting time and computing cost, and may even cause the model to overthink simple questions. To address this problem, the paper proposes SmartThinker. Its main idea is that a model should not always think more, nor should it always be forced to think less. Instead, it should learn to use the right amount of reasoning based on how difficult each question is. For easier questions, SmartThinker removes unnecessary steps. For harder questions, it keeps the useful reasoning needed to reach the correct answer. It also avoids treating long but correct answers as bad examples simply because they are long. Experiments show that SmartThinker can shorten model responses while maintaining or improving accuracy across several math tests and model sizes. In some cases, it reduces the average response length by about half and improves performance on harder tasks. Overall, the study shows that teaching models to “think just enough” is more effective than simply making them think more or less.