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

Chenzhi Hu; Qinzhe Hu; Yuhang Xu; Junyi Chen; Ruijie Wang; Shengzhong Liu; Jianxin Li; Fan Wu

cs.CLMar 9, 2026

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

Authors: Chenzhi Hu, Qinzhe Hu, Yuhang Xu, Junyi Chen, Ruijie Wang, Shengzhong Liu, Jianxin Li, Fan Wu, +1 more

Organizations: Shanghai Jiao Tong University · Beihang University

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.

Figures & tables

Figure 1: An illustrative comparison between the base model and SmartThinker on an AIME25 problem.

Figure 2: Overview of SmartThinker ’s dynamic reward and advantage calculation process.

Figure 3: Estimation of the optimal length. The subfigure above shows the distribution of all samples and correct samples and the one below shows the span with the highest accuracy and the estimated optimal length.

Method	#RL Steps ↓	Math500		AIME25		AMC23		Average
Method	#RL Steps ↓	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len. ↓	Acc. (%) ↑	AE ↑
DeepSeek-R1-Distill-Qwen-1.5B
Base Model	N/A	5420	84.9	15199	24.2	9320	73.1	9980	60.7	N/A
Truncated Think 4k	N/A	2950	71.4	4299	11.7	3627	42.5	3625	41.9	-0.91
Truncated Think 8k	N/A	3971	72.0	7531	17.5	5471	50.6	5658	46.7	-0.72
ShorterBetter	300	1008	71.0	3727	19.0	2246	66.9	2327	52.3	0.07

Table 1: Performance comparison across base models and benchmarks. The SmartThinker rows highlight our results.

Method	#RL Steps ↓	Math500		AIME25		AMC23		Average
Method	#RL Steps ↓	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len. ↓	Acc. (%) ↑	AE ↑
Base Model	N/A	5420	84.9	15199	24.2	9320	73.1	9980	60.7	N/A
AutoThink + SmartThinker
AutoThink-Stage2	440	3593	85.4	10571	27.5	6206	75.0	6720	62.6	0.41
AutoThink-Stage2 → Stage3	440 + 60	2196	85.2	9084	24.2	4117	73.8	5132	61.1	0.50
AutoThink-Stage2 → SmartThinker	440 + 50	2792	84.9	8760	29.2	5134	74.4	5562	62.8	0.55

Table 2: The result of combining AutoThink-Stage2 and SmartThinker .

DeepSeek-R1-Distill-Qwen-1.5B
Models	MMLU		MathQA		LiveCode		HumanEval		Average
Models	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len.↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑	Len. ↓	Acc.(%) ↑
Base Model	1686	46.02	5206	84	10668	24.2	4740	68.9	5575	55.78
SmartThinker	994	46.48	2581	84.6	7658	25.4	3099	69.5	3583	56.495
Qwen3-4B-Thinking-2507
Base Model	2978	80.8	3253	91.3	12388	68.1	4507	95.7	5781.5	83.975

Table 3: OOD evaluation on models of different scales.

Figure 4: Changes in various metrics during the training process

Method	Math500		AIME25		AMC23		Average
Method	len	acc	len	acc	len	acc	len	acc
Fixed Coefficient	1897	80.3	6049	21.7	2987	70.6	3644	57.5
Symmetric	2902	83.4	8761	24.2	4927	72.9	5530	60.2
Linear	2439	83	6909	21.7	3379	70	4242	58.2
SmartThinker	2645	84.5	8431	25	4421	76.3	5169	61.9

Table 4: Ablation study on DeepSeek-R1-Distill-Qwen-1.5B

Category	Base Model	SmartThinker
Pivotal Reasoning	6.13%	16.04%
Repetition & Rephrasing	1.13%	0.68%
Exploring Alternatives	86.86%	77.29%
Verification & Explanation	2.73%	2.90%
Non-Substantive Statement	3.05%	3.09%

Table 5: SmartThinker ’s impact on CoT structure.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: The curve of

h(l)

under different conditions.

Condition	Mathematical Property	Reasoning Intuition	$l^{\text{opt}}$ Strategy
$\sigma_{1}^{2}>\sigma_{2}^{2}$	Local Maximum	Existence of a ”Golden Length”	Computed $l^{\text{opt}}$
$\sigma_{1}^{2}=\sigma_{2}^{2},\mu_{1}>\mu_{2}$	Linear Decrease	Shorter is strictly better	$0$ (min length)
$\sigma_{1}^{2}=\sigma_{2}^{2},\mu_{1}<\mu_{2}$	Linear Increase	Complexity aids accuracy	$+\infty$ (max length)
$\sigma_{1}^{2}<\sigma_{2}^{2}$	Local Minimum	Unstable middle-ground	0 (boundary clipping)

Appendix

Table 6: Behavior of

l^{\text{opt}}

under different

\mu

and

\sigma

conditions.

	$P_{\text{SW}}$	$P_{\text{KS}}$ and $P_{\text{shape}}$	$P_{\text{KS}}$
All Responses	20.0	52.3	62.9
Correct Responses	27.7	60.0	74.3

Appendix

Table 7: Gaussian test of length distributions.

	Model	Pass Rate
All Responses	DeepSeek-R1-Distill-1.5B	89.90%
All Responses	SmarThinker-Distill-1.5B	94.50%
Correct Responses	DeepSeek-R1-Distill-1.5B	98.00%
Correct Responses	SmarThinker-Distill-1.5B	100%

Appendix

Table 8: Unimodal test results

Figure 6: Comparison of optimal length under different pass counts.

Figure 7: Test results of baseline and SmartThinker on four out-of-domain benchmarks.

Explore similar work

To Think or Not to Think: Allocating Reasoning Where It Helps
Aug 30, 2026Zhengdong He, Yunfan Zhou, Jianguo Yao +2Efficient Language Model Reasoning RL for Language Model Reasoning
Shanghai Key Laboratory of Scalable Computing and Systems · School of Computer Science, Shanghai Jiao Tong University
Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs
Feb 28, 2026Jie Cao, Tianwei Lin, Juekai Lin +7Efficient Language Model Reasoning CoT Reasoning
Zhejiang University · Tencent
OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping
Jul 13, 2026Mohammed Ehab, Aymane El Gadarri, Vivek F. Farias +2Overthinking in Language Models Sequential Decision Making
MIT · Columbia Business School

Aug 30, 2026J/K move · Enter open · S save

To Think or Not to Think: Allocating Reasoning Where It Helps

Zhengdong He, Yunfan Zhou, Jianguo Yao, Haibing Guan+1

Shanghai Key Laboratory of Scalable Computing and Systems · School of Computer Science, Shanghai Jiao Tong University

Efficient Language Model Reasoning RL for Language Model Reasoning

Open paper AskPDF arXiv

Reinforcement learning (RL) has proven effective in enhancing the reasoning performance of large language models (LLMs), particularly in complex mathematical and programming tasks. However, this capability comes with systematic length misallocation, in which models devote excessive reasoning to simple questions while terminating prematurely on harder ones, degrading inference efficiency with negligible accuracy improvement. Many length-adaptive methods mitigate this issue by allocating token budgets according to question difficulty, under the implicit assumption that harder questions benefit monotonically from extended reasoning. In contrast, we find that the effect of reasoning length on accuracy is concentrated on partially solvable questions. Our further analysis reveals that explicit length rewards can produce unintended training dynamics. Motivated by these findings, we propose CARE---Contrastive Accuracy Reward Estimation---which compares the beneficial length adjustment per question from online sampled responses and applies adaptive length rewards within Group Relative Policy Optimization, with no extra hyperparameters or additional inference cost. Experiments across multiple reasoning benchmarks demonstrate that our method improves Pass@1 by up to

4\%

while simultaneously reducing reasoning length by

37\%

, achieving higher token efficiency. Code will be available upon the acceptance of this paper.

Abstract

Explore similar work

To Think or Not to Think: Allocating Reasoning Where It Helps

Draft-Thinking: Learning Efficient Reasoning in Long Chain-of-Thought LLMs

OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping