mlsanityData to 14 Oct 2026

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

arXiv 2603.17449 ↗First posted 18 Mar 2026cs.CLcited by 2

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

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.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 48 of its 61 references that are in the map. First posted on arXiv 18 Mar 2026.

Builds on

48 of 61 references.

Closest ancestors · 33

Same direction, newest first

  1. 2026ConPress: Learning Efficient Reasoning from Multi-Question Contextual Pressure1 Feb
  2. 2026Anti-Length Shift: Dynamic Outlier Truncation for Training Efficient Reasoning Models7 Jan
  3. 2025Entropy-Guided Reasoning Compression18 Nov
  4. 2025Confidence-Guided Stepwise Model Routing for Cost-Efficient Reasoning9 Nov
  5. 2025Efficient Reasoning via Thought-Training and Thought-Free Inference5 Nov
  6. 2025DART: Difficulty-Adaptive Reasoning Truncation for Efficient Large Language Models3 Nov
  7. 2025SpecExit: Accelerating Large Reasoning Model via Speculative Exit29 Sep
  8. 2025Your Models Have Thought Enough: Training Large Reasoning Models to Stop Overthinking27 Sep
  9. 2025Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression7 Aug
  10. 2025Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy5 Aug
  11. 2025MUR: Momentum Uncertainty guided Reasoning for Large Language Models20 Jul
  12. 2025Optimizing Length Compression in Large Reasoning Models17 Jun
21 more
  1. 2025AlphaOne: Reasoning Models Thinking Slow and Fast at Test Time30 May
  2. 2025The Price of a Second Thought: On the Evaluation of Reasoning Efficiency in Large Language Models28 May
  3. 2025AdaCtrl: Towards Adaptive and Controllable Reasoning via Difficulty-Aware Budgeting24 May
  4. 2025Learn to Reason Efficiently with Adaptive Length-based Reward Shaping21 May
  5. 2025AdaptThink: Reasoning Models Can Learn When to Think19 May
  6. 2025Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RL16 May
  7. 2025ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning30 Apr
  8. 2025Dynamic Early Exit in Reasoning Models22 Apr
  9. 2025Reasoning Models Can Be Effective Without Thinking14 Apr
  10. 2025Retro-Search: Exploring Untaken Paths for Deeper and Efficient Reasoning6 Apr
  11. 2025ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning2 Apr
  12. 2025L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning6 Mar
  13. 2025DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models6 Mar
  14. 2025How Well do LLMs Compress Their Own Chain-of-Thought? A Token Complexity Approach3 Mar
  15. 2025Self-Training Elicits Concise Reasoning in Large Language Models27 Feb
  16. 2025LightThinker: Thinking Step-by-Step Compression21 Feb
  17. 2025TokenSkip: Controllable Chain-of-Thought Compression in LLMs17 Feb
  18. 2025CoT-Valve: Length-Compressible Chain-of-Thought Tuning13 Feb
  19. 2025LLM Pretraining with Continuous Concepts12 Feb
  20. 2025Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning5 Feb
  21. 2025O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning22 Jan

Other references · 15

  1. 2025Qwen3 Technical Reportcited by 8,820
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  3. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  4. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  5. 2021Measuring Mathematical Problem Solving With the MATH Datasetcited by 5,371
  6. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 2,057
  7. 2024OpenAI o1 System Cardcited by 1,244
  8. 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 982
  9. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  10. 2022Complexity-Based Prompting for Multi-Step Reasoningcited by 388
  11. 2025Chain of Draft: Thinking Faster by Writing Lesscited by 297
  12. 2025ARM: Adaptive Reasoning Modelcited by 39
3 more
  1. 2025DRP: Distilled Reasoning Pruning with Skill-aware Step Decomposition for Efficient Large Reasoning Modelscited by 30
  2. 2025Concise Reasoning, Big Gains: Pruning Long Reasoning Trace with Difficulty-Aware Promptingcited by 15
  3. 2025ConciseRL: Conciseness-Guided Reinforcement Learning for Efficient Reasoning Modelscited by 14

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Uncovering the Representation Geometry of Minimal Cores in Overcomplete Reasoning Tracescited by 1
  2. 2026CLORE: Content-Level Optimization for Reasoning Efficiencycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026CtrlCoT: Dual-Granularity Chain-of-Thought Compression for Controllable Reasoning
  2. 2025R1-Compress: Long Chain-of-Thought Compression via Chunk Compression and Search
  3. 2026PACE: Prefix-Protected and Difficulty-Aware Compression for Efficient Reasoning
  4. 2025Breadcrumbs Reasoning: Memory-Efficient Reasoning with Compression Beacons
  5. 2025Pruning the Unsurprising: Efficient LLM Reasoning via First-Token Surprisal
  6. 2025TrimR: Verifier-based Training-Free Thinking Compression for Efficient Test-Time Scaling
  7. 2026Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression
  8. 2025ConCISE: Confidence-guided Compression in Step-by-step Efficient Reasoning