TRiMS: Real-Time Tracking of Minimal Sufficient Length for Efficient Reasoning via RL

Tingcheng Bian; Jinchang Luo; Mingquan Cheng; Jinyu Zhang; Xiaoling Xia; Ni Li; Yan Tao; Haiwei Wang

cs.CLMar 18, 2026

TRiMS: Real-Time Tracking of Minimal Sufficient Length for Efficient Reasoning via RL

Authors: Tingcheng Bian, Jinchang Luo, Mingquan Cheng, Jinyu Zhang, Xiaoling Xia, Ni Li, Yan Tao, Haiwei Wang

Organizations: Baidu Inc. · Shenzhen University · Harbin Institute of Technology

Abstract

Large language models achieve breakthroughs in complex reasoning via long chain-of-thought sequences. However, this often leads to severe reasoning inflation, causing substantial computational redundancy. To maximize Intelligence per Token, we introduce a theoretical metric, MSL-Minimal Sufficient Length. MSL rigorously characterizes the shortest reasoning length that preserves answer correctness. We provide a recursive definition based on independently sampled sequences and prove the existence of its limit, establishing the first measurable lower bound for reasoning-chain compression. Building on an analysis of mainstream CoT compression strategies, we identify key structural factors enabling a model to approach MSL. Based on these insights, we propose TRiMS which employs the GRPO algorithm in conjunction with MSL-based estimation during training, while mitigating instabilities during the training process through dynamic batch aggregation and advantage computation using batch-level standard deviation. TRiMS achieves over 80% CoT token reduction with a minor accuracy boost across all benchmarks.

Figures & tables

Panel of Figure 1 — Figure 1: (Top) Conceptual illustration of diverse reasoning paths generated by LLM. (Bottom) Expected token length of the shortest correct reasoning path(SCPT@K) as a function of the sampling times $k$ .

Panel of Figure 2 — Figure 2: Expected token length of the shortest correct reasoning path versus sample number $k$ . (a–d) Different sampling strategies. (e) Model scaling under varying difficulty levels.

Figure 3: Overview of the TRiMS framework

Figure 4: Percentage of degenerate groups across training steps

Figure 5: Proportion of correct answers distribution across different length thresholds and difficulty levels.

DeepSeek-R1-Distill-Qwen-1.5B
Method	AMC23		AIME24		MATH-500		Minerva		Olympiad		Avg
Method	Acc	Token	Acc	Token	Acc	Token	Acc	Token	Acc	Token	$\Delta$ Acc (%)	$\Delta$ Token (%)	IPT
Vanilla	61.1	8458.9	29.2	12303.4	82.3	4619.2	25.7	6284.1	41.0	8994.3	–	–	7.2
LC-R1	56.5	3862.9	19.3	6737.5	71.2	2080.4	26.5	3827.3	37.9	3950.1	-11.9	-49.9	13.6
Laser	60.2	4714.9	30.4	5713.1	84.1	1974.1	33.1	3218.3	44.3	5645.9	+8.4	-48.2	15.8
AutoThink	57.8	4244.9	30.0	8984.4	85.6	2111.5	29.0	3031.5	44.7	5072.6	+1.9	-45.3	14.9

Table 1: Comparison of PASS@1 accuracy, average output length (Token), and Intelligence Per Token (IPT) between the TRiMS model and baseline models across all benchmark datasets.

Figure 6: Distribution of correct responses per group (

k

) under different truncation strategies (

G=16

Panel of Figure 7 — Figure 7: Group-level vs. batch-level advantage normalization throughout training.

Figure 8: Per-problem generation lengths on AIME24. TRiMS (blue dots) is compared against the empirically estimated MSL (black curve, SCPT@k with k=512), sorted by descending MSL estimate.

Figure 9: Reasoning trace statistics on AIME-24: Comparison of reasoning steps and overthinking keyword frequencies.

Figure 10: Average response length of TRiMS across five difficulty levels on MATH-500 (1.5B model).

Figure 11: TRiMS fits more parallel trajectories within the same budget, substantially outperforming the Vanilla model on AMC23.

Figure 12: Qwen3-4B Tokens before and after training.

DeepSeek-R1-Distill-Qwen-1.5B
Method	MMLU
Method	Acc	Token	$\Delta$ Acc	$\Delta$ Token	IPT
Vanilla	38.1	1718.7	–	–	22.2
TRiMS	48.2	388.1	+26.5	-77.4	124.2
DeepSeek-R1-Distill-Qwen-7B
Vanilla	63.7	1158.5	–	–	55.0

Table 2: OOD generalization results on MMLU.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration	Temperature	top- $p$	top- $k$	Default
Low-temperature	0.3	0.95	$-1$	No
Moderate-temperature	0.6	0.95	$-1$	Yes
High-temperature	1.0	0.95	$-1$	No
Constrained Stochastic	0.6	1.0	100	No

Appendix

Table 3: Sampling configurations used throughout the paper. The Moderate-temperature setting is adopted as the default unless otherwise specified.

\mathcal{C}_{i}=\{L_{j}\mid\text{sample }j\text{ is correct}\},\quad c_{i}=|\mathcal{C}_{i}|

Appendix

Algorithm 1 Unbiased Estimation of Shortest Correct Path Tokens (SCPT@

k

)

Method	1.0	2.0	3.0	4.0	5.0
Vanilla	0.9070	0.9333	0.8952	0.7969	0.6567
TRiMS	0.9535	0.9111	0.9429	0.8125	0.6493

Appendix

Table 4: Accuracy comparison between Vanilla and TRiMS across difficulty levels.

Qwen3-4B
Method	AMC23		AIME24		MATH-500		Minerva		Olympiad		Avg
Method	Acc	Token	Acc	Token	Acc	Token	Acc	Token	Acc	Token	$\Delta$ Acc (%)	$\Delta$ Token (%)	IPT
Vanilla	77.1	9060.0	57.9	11933.0	91.6	5021.6	41.5	6673.1	58.2	9243.0	–	–	8.8
TRiMS	80.7	1046.2	53.5	1713.6	91.0	662.2	41.9	783.0	57.9	1081.6	-0.63	-88.3	70.6

Appendix

Table 5: Comparison of PASS@1 accuracy, average output length (Token), and Intelligence Per Token (IPT) between the TRiMS model and baseline models across all benchmark datasets.

Figure 13: PASS@

k

accuracy on AMC23 (DeepSeek-R1-Distill-Qwen-1.5B) as a function of the number of samples

k

. The shaded region highlights the accuracy gap between TRiMS and the Vanilla baseline.

Figure 14: Expected token length of the shortest correct reasoning path for Qwen3-4B across varying difficulty levels as a function of sampling attempts

k

Panel of Figure 15 — Figure 15: Case study on an AIME problem. Vanilla uses 4452 tokens with extensive step-by-step reasoning, while TRiMS reaches the same correct answer using only 995 tokens.

Panel of Figure 16 — Figure 16: Case study on a continued fraction problem. Vanilla uses 3083 tokens with repeated verification, while TRiMS solves it in only 592 tokens.

Panel of Figure 17 — Figure 17: Case study on a coordinate conversion problem. Vanilla uses 1555 tokens with extensive self-verification, while TRiMS solves it in only 413 tokens.

Panel of Figure 18 — Figure 18: Case study on a geometry problem. Vanilla exhausts the 16384 token limit without reaching an answer, while TRiMS correctly solves it in 3121 tokens.

Efficient Language Model Reasoning

Efficient language model reasoning studies how to improve problem-solving accuracy while reducing inference, memory, latency, and training costs.

See all 371 papers

ExpThink: Experience-Guided Reinforcement Learning for Adaptive Chain-of-Thought Compression
May 8, 2026Tingcheng Bian, Yuzhe Zhang, Jing Jin +5Reward Shaping RL for Language Model Reasoning
1Baidu Inc. · 2Shenzhen University · 3Peking University +2
CRISP: Compressing Redundancy in Chain-of-Thought via Intrinsic Saliency Pruning
Apr 19, 2026Yangsong Lan, Hongliang Dai, Piji LiLLM Pruning LLM Compression
College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, Nanjing, China · The Key Laboratory of Brain-Machine Intelligence Technology, Ministry of Education, Nanjing, China.
SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning
May 29, 2026Jian Yao, Xiongcai Luo, Ran Cheng +1RL for Language Model Reasoning Efficient Language Model Inference
Department of Data Science and Artificial Intelligence, The Hong Kong Polytechnic University · Bytedance · The Hong Kong Polytechnic University-Daya Bay Technology and Innovation Research Institute, Huizhou,Guangdong Province, China.

May 8, 2026J/K move · Enter open · S save

ExpThink: Experience-Guided Reinforcement Learning for Adaptive Chain-of-Thought Compression

Tingcheng Bian, Yuzhe Zhang, Jing Jin, Jinchang Luo+4

1Baidu Inc. · 2Shenzhen University · 3Peking University+2

Reward Shaping RL for Language Model Reasoning Efficient Language Model Reasoning Reward Design for RL CoT Compression Large Reasoning Models

Open paper AskPDF arXiv

Large reasoning models (LRMs) achieve strong performance via extended chain-of-thought (CoT) reasoning, yet suffer from excessive token consumption and high inference latency. Existing reinforcement learning (RL) approaches for CoT compression rely on uniform, static length penalties that neglect model capability dynamics and problem-level difficulty variation. We propose ExpThink\xspace, an RL framework that addresses both dimensions through two complementary mechanisms. First, experience-guided reward shaping tracks the shortest correct solution found so far for each problem and applies a three-tier reward: full credit for concise correct responses, discounted credit for verbose correct ones, and zero for incorrect ones. The threshold tightens automatically with model improvement, forming a self-evolving curriculum that requires no manual scheduling. Second, difficulty-adaptive advantage replaces standard deviation normalization with correct-count normalization, yielding monotonically difficulty-scaled gradients that amplify learning on hard problems to preserve accuracy while suppressing gradients on easy ones to encourage brevity. Together, these mechanisms enforce an accuracy-first, compression-second training objective. Experiments on multiple mathematical reasoning benchmarks demonstrate that ExpThink\xspace reduces average response length by up to 77% while simultaneously improving accuracy, achieving up to

3\times

higher accuracy-efficiency ratio (accuracy divided by average token count) than the vanilla baseline and outperforming existing RL-based compression methods on both metrics.

Abstract

Efficient Language Model Reasoning

Explore similar work

ExpThink: Experience-Guided Reinforcement Learning for Adaptive Chain-of-Thought Compression

CRISP: Compressing Redundancy in Chain-of-Thought via Intrinsic Saliency Pruning

SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning