mlsanityData to 14 Oct 2026

Entropy After </Think> for reasoning model early exiting

arXiv 2509.26522 ↗First posted 30 Sep 2025cs.LGcited by 24

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

also in 1 more
Abstract (arXiv)

Reasoning LLMs show improved performance with longer chains of thought. However, recent work has highlighted their tendency to overthink, continuing to revise answers even after reaching the correct solution. We quantitatively confirm this inefficiency from the distribution dynamics perspective by tracking Pass@1 for answers averaged over a large number of rollouts and find the model often begins to always produce the correct answer early in the reasoning, making extra reasoning tokens wasteful. To detect and prevent overthinking, we propose a simple and inexpensive novel signal, Entropy After </Think> (EAT), for monitoring and deciding whether to exit reasoning early. By appending a stop thinking token (</think>) and monitoring the entropy of the following token as the model reasons, we obtain a trajectory that decreases and stabilizes when Pass@1 plateaus; thresholding its variance under an exponential moving average yields a practical stopping rule. Importantly, our approach enables adaptively allocating compute based on the EAT trajectory, allowing us to spend compute in a more efficient way compared with fixing the token budget for all questions. Empirically, on MATH500 and AIME2025, EAT reduces token usage by 12 - 22% without harming accuracy. EAT also remains effective in black box settings where logits from the reasoning model are not accessible, and EAT is computed with proxy models: We verified the feasibility via early stopping Llama 70B with a 1.5B model and Claude 3.7 with a local 4B model.

Key facts. Cited by 24 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 32 of its 46 references that are in the map. First posted on arXiv 30 Sep 2025.

Builds on

32 of 46 references.

Closest ancestors · 9

Same direction, newest first

  1. 2025Think or Not? Exploring Thinking Efficiency in Large Reasoning Models via an Information-Theoretic Lens23 May
  2. 2025Optimizing Anytime Reasoning via Budget Relative Policy Optimization19 May
  3. 2025Dynamic Early Exit in Reasoning Models22 Apr
  4. 2025Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification7 Apr
  5. 2025Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models20 Mar
  6. 2025Is That Your Final Answer? Test-Time Scaling Improves Selective Question Answering19 Feb
  7. 2025s1: Simple test-time scaling31 Jan
  8. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs30 Dec
  9. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters6 Aug

Other references · 23

  1. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  2. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  3. 2023Let's Verify Step by Stepcited by 3,789
  4. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  5. 2024Qwen2 Technical Reportcited by 2,120
  6. 2017BranchyNet: Fast Inference via Early Exiting from Deep Neural Networkscited by 596
  7. 2025Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoningcited by 559
  8. 2022Efficiently Scaling Transformer Inferencecited by 539
  9. 2016Adaptive Computation Time for Recurrent Neural Networkscited by 506
  10. 2022Confident Adaptive Language Modelingcited by 310
  11. 2025Reasoning with Exploration: An Entropy Perspectivecited by 281
  12. 2019Depth-Adaptive Transformercited by 236
11 more
  1. 2019Rank-normalization, folding, and localization: An improved widehatR for assessing convergence of MCMCcited by 93
  2. 2025S-GRPO: Early Exit via Reinforcement Learning in Reasoning Modelscited by 87
  3. 2025Jointly Reinforcing Diversity and Quality in Language Model Generationscited by 80
  4. 2025Sample More to Think Less: Group Filtered Policy Optimization for Concise Reasoningcited by 75
  5. 2025Answer Matching Outperforms Multiple Choice for Language Model Evaluationcited by 41
  6. 2025Answer Convergence as a Signal for Early Stopping in Reasoningcited by 40
  7. 2024Understanding Chain-of-Thought in LLMs through Information Theorycited by 37
  8. 2024Fast yet Safe: Early-Exiting with Risk Controlcited by 30
  9. 2025LLM Probability Concentration: How Alignment Shrinks the Generative Horizoncited by 26
  10. 2025RLP: Reinforcement as a Pretraining Objectivecited by 25
  11. 2025Early Stopping Chain-of-thoughts in Large Language Modelscited by 20

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 4; 2026: 20.

24 papers.

Sort:

By area:

  1. 2026Parallel-Probe: Towards Efficient Parallel Thinking via 2D Probingcited by 14
  2. 2026Are Your Reasoning Models Reasoning or Guessing? A Mechanistic Analysis of Hierarchical Reasoning Modelscited by 10
  3. 2026Conformal Thinking: Risk Control for Reasoning on a Compute Budgetcited by 10
  4. 2026LLMs Improving LLMs: Agentic Discovery for Test-Time Scalingcited by 10
  5. 2026Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Modelscited by 7
  6. 2026TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoningcited by 6
  7. 2026Early Stopping for Large Reasoning Models via Confidence Dynamicscited by 6
  8. 2025PRInTS: Reward Modeling for Long-Horizon Information Seekingcited by 5
  9. 2026On-Policy Supervised Fine-Tuning for Efficient Reasoningcited by 4
  10. 2026Unveiling the Entropy Dynamics of Chain-of-Thought Reasoningcited by 4
  11. 2025Controlling the Risk of Corrupted Contexts for Language Models via Early-Exitingcited by 3
  12. 2026interwhen: A Generalizable Framework for Steering Reasoning Models with Test-time Verificationcited by 2
12 more
  1. 2026The Potential of CoT for Reasoning: A Closer Look at Trace Dynamicscited by 2
  2. 2026When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Modelscited by 2
  3. 2026Reasoning as Compression: Unifying Budget Forcing via the Conditional Information Bottleneckcited by 1
  4. 2026ROM: Real-time Overthinking Mitigation via Streaming Detection and Interventioncited by 1
  5. 2026Think Twice Before You Write -- an Entropy-based Decoding Strategy to Enhance LLM Reasoningcited by 1
  6. 2026DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?cited by 1
  7. 2026SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probingcited by 1
  8. 2025LiteStage: Latency-aware Layer Skipping for Multi-stage Reasoningcited by 0
  9. 2025Step-Tagging: Toward controlling the generation of Language Reasoning Models through step monitoringcited by 0
  10. 2026TRACES: Tagging Reasoning Steps for Adaptive Cost-Efficient Early-Stoppingcited by 0
  11. 2026Rewarding Efficient Reasoning Improves Abstention on Underspecified Tasks in Reasoning Modelscited by 0
  12. 2026Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiencycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Explore Briefly, Then Decide: Mitigating LLM Overthinking via Cumulative Entropy Regulation
  2. 2025Think Just Enough: Sequence-Level Entropy as a Confidence Signal for LLM Reasoning
  3. 2025Measuring Reasoning Utility in LLMs via Conditional Entropy Reduction
  4. 2026When Do LLMs Reason? A Dynamical Systems View via Entropy Phase Transitions
  5. 2026Answer-Distribution Trajectories: A Stochastic-Dynamics View of LLM Reasoning
  6. 2026Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring
  7. 2026Thinking Past the Answer: Evaluating Harmful Overthinking in Large Reasoning Models
  8. 2025The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis