Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models

• URL: http://arxiv.org/abs/2605.17672v1
• Date: Sun, 17 May 2026 22:04:11 GMT
• Title: Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models
• Abstract summary: Large Reasoning Models (LRMs) achieve strong performance by generating long chains of thought (CoT)<n>LRMs often overthink, continuing to reason after a solution has already stabilized and thereby wasting tokens and increasing latency.<n>We propose PUMA, a plug-and-play framework that combines a lightweight Redundancy Detector with answer-level verification.
• Score: 17.588873255093596
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large Reasoning Models (LRMs) achieve strong performance by generating long chains of thought (CoT), but often overthink, continuing to reason after a solution has already stabilized and thereby wasting tokens and increasing latency. Existing inference-time early-exit methods rely primarily on answer-level signals, such as confidence or trial-answer consistency, to decide when to stop. However, these signals mainly reflect answer readiness rather than reasoning convergence: they may trigger before the model has finished exploring or self-correcting, causing premature exits that can degrade final-answer accuracy and leave the retained reasoning chain semantically incomplete. We identify reasoning-level semantic redundancy as a complementary signal for semantic-preserving early exit: when successive steps no longer add novel progress and instead revisit established conclusions, the reasoning trajectory has likely converged. Building on this insight, we propose PUMA, a plug-and-play framework that combines a lightweight Redundancy Detector with answer-level verification. The detector flags semantically redundant candidate exits, while verification confirms whether stopping is safe, allowing PUMA to remove redundant continuation while preserving both answer accuracy and a coherent reasoning prefix. Across five LRMs and five challenging reasoning benchmarks, PUMA achieves 26.2% average token reduction while preserving accuracy and retained CoT quality. Additional experiments on code generation, zero-shot vision-language reasoning, and learned stopping-policy internalization further demonstrate that reasoning-level redundancy is a robust, transferable, and learnable signal for efficient reasoning. Our code is available at \url{https://github.com/giovanni-vaccarino/PUMA}.

Related papers

• </think> Doesn't Stop Reasoning: Analysis of Spurious CoT Termination [8.097661521171581]
  We study a strategy that injects an end-of-think token (EoT) at this point to trigger the reasoning-to-answering transition.<n>We find that the injected EoT does not always induce a clean answering phase.<n>We call this spurious CoT termination, where reasoning-like generation continues into the answering phase.
  arXiv  Detail & Related papers  (2026-09-03T10:25:23Z)
• SABER: Stability-Aware Early Exit for LLM Reasoning via Adversarial Branch Probing [32.867599324353456]
  SABER is a training-free framework for stability-aware early exit via adversarial branch probing.<n>It reduces reasoning token consumption by 30.2%--39.8% on average while maintaining competitive accuracy with full-length reasoning.
  arXiv  Detail & Related papers  (2026-08-28T06:13:23Z)
• Understanding and Mitigating Premature Confidence for Better LLM Reasoning [76.16007941549857]
  Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps.<n>We find such a signal in how the model's confidence evolves during reasoning.<n>We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early.
  arXiv  Detail & Related papers  (2026-05-23T04:42:45Z)
• CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning [22.944748148277146]
  CopT is a reformulated reasoning pipeline that reverses the usual order of thinking and answering.<n>CopT elicits a draft answer and then invokes subsequent on-policy thinking conditioned on its own draft answer.<n>CopT improves peak accuracy by up to 23% and reduces token usage by up to 57% at comparable or higher accuracy.
  arXiv  Detail & Related papers  (2026-05-19T16:28:53Z)
• Efficient Test-Time Scaling via Temporal Reasoning Aggregation [5.87118425760756]
  We propose TRACE, a training-free framework for efficient test-time scaling.<n>TRACE determines when to terminate reasoning based on temporal aggregation of multi-step evidence.<n>We show that TRACE reduces reasoning token usage by 25-30% on average while maintaining accuracy within 1-2% of full-length reasoning.
  arXiv  Detail & Related papers  (2026-04-19T07:39:40Z)
• Early Stopping for Large Reasoning Models via Confidence Dynamics [55.67938134245981]
  Large reasoning models rely on long chain-of-thought generation to solve complex problems.<n>Key challenge is determining when the model should stop reasoning and produce the final answer.<n>We propose CoDE-Stop, an early stopping method that leverages the dynamics of intermediate answer confidence to decide when to terminate reasoning.
  arXiv  Detail & Related papers  (2026-04-06T17:59:45Z)
• SPOT: Span-level Pause-of-Thought for Efficient and Interpretable Latent Reasoning in Large Language Models [15.95627037350657]
  Explicit Chain-of-Thought (CoT) computation incurs high inference cost due to internalize token-level traces.<n>We propose SPOT, a flexible framework that compresses explicit CoT into compact latent pause tokens without enforcing a fixed response template.<n>Experiments on reasoning benchmarks demonstrate that SPOT improves accuracy by 2.3 points on average while reducing generated tokens by 37.5%.
  arXiv  Detail & Related papers  (2026-03-06T12:34:27Z)
• Balancing Faithfulness and Performance in Reasoning via Multi-Listener Soft Execution [79.98699884805636]
  Reasoning Execution by Multiple Listeners (REMUL) is a multi-party reinforcement learning approach.<n>REMUL builds on the hypothesis that reasoning traces which other parties can follow will be more faithful.<n>Speakers are rewarded for producing reasoning that is clear to listeners.
  arXiv  Detail & Related papers  (2026-02-18T02:55:55Z)
• APR: Penalizing Structural Redundancy in Large Reasoning Models via Anchor-based Process Rewards [61.52322047892064]
  Test-Time Scaling (TTS) has significantly enhanced the capabilities of Large Reasoning Models (LRMs)<n>We observe that LRMs frequently conduct repetitive self-verification without revision even after obtaining the final answer during the reasoning process.<n>We propose Anchor-based Process Reward (APR), a structure-aware reward shaping method that localizes the reasoning anchor and penalizes exclusively the post-anchor AST.
  arXiv  Detail & Related papers  (2026-01-31T14:53:20Z)
• Thinking Traps in Long Chain-of-Thought: A Measurable Study and Trap-Aware Adaptive Restart [27.904791075662896]
  We introduce TAAR (Trap-Aware Adaptive Restart), a test-time control framework that trains a diagnostic policy to predict two signals from partial trajectories.<n>At inference time, TAAR truncates the trajectory before the predicted trap segment and adaptively restarts decoding.<n>Experiments show that TAAR improves reasoning performance without fine-tuning base model parameters.
  arXiv  Detail & Related papers  (2026-01-17T07:26:02Z)
• Stop Spinning Wheels: Mitigating LLM Overthinking via Mining Patterns for Early Reasoning Exit [114.83867400179354]
  Overthinking can degrade overall performance of large language models.<n>We categorize reasoning into three stages: insufficient exploration stage, compensatory reasoning stage, and reasoning convergence stage.<n>We develop a lightweight thresholding strategy based on rules to improve reasoning accuracy.
  arXiv  Detail & Related papers  (2025-08-25T03:17:17Z)
• ThinkLess: A Training-Free Inference-Efficient Method for Reducing Reasoning Redundancy [8.962703809086628]
  ThinkLess is an inference-efficient framework that terminates reasoning generation early and maintains output quality without modifying the model.<n>We show that ThinkLess achieves comparable accuracy to full-length Chain-of-Thought (CoT) decoding while greatly reducing decoding time and memory consumption.
  arXiv  Detail & Related papers  (2025-05-21T15:58:16Z)
• Mitigating Misleading Chain-of-Thought Reasoning with Selective Filtering [59.495717939664246]
  Large language models have manifested remarkable capabilities by leveraging chain-of-thought (CoT) reasoning techniques to solve intricate questions. We propose a novel approach called the selective filtering reasoner (SelF-Reasoner) that assesses the entailment relationship between the question and the candidate reasoning chain. SelF-Reasoner improves the fine-tuned T5 baseline consistently over the ScienceQA, ECQA, and LastLetter tasks.
  arXiv  Detail & Related papers  (2024-03-28T06:28:35Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.