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

Hugging Face Daily Papers Papers
reasoning-models early-exit semantic-redundancy inference-efficiency chain-of-thought token-reduction

Summary

This paper introduces PUMA, a plug-and-play framework that detects semantic redundancy in chain-of-thought reasoning to enable early exit, achieving 26.2% average token reduction across multiple models and benchmarks while preserving accuracy and reasoning quality.

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 https://github.com/giovanni-vaccarino/PUMA.
Original Article
View Cached Full Text

Cached at: 05/19/26, 06:31 AM

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

Source: https://huggingface.co/papers/2605.17672

Abstract

Researchers introduce PUMA, a framework that uses semantic redundancy detection to improve reasoning efficiency in large models by identifying when continued thinking provides no new insights, thus reducing computational waste while maintaining answer accuracy and reasoning quality.

Large Reasoning Models(LRMs) achieve strong performance by generating longchains of thought(CoT), but often overthink, continuing to reason after a solution has already stabilized and thereby wasting tokens and increasing latency. Existing inference-timeearly-exit methodsrely primarily onanswer-level signals, such as confidence or trial-answer consistency, to decide when to stop. However, these signals mainly reflect answer readiness rather thanreasoning 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-levelsemantic redundancyas 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 lightweightRedundancy Detectorwithanswer-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% averagetoken reductionwhile 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 https://github.com/giovanni-vaccarino/PUMA.

View arXiv pageView PDFGitHub2Add to collection

Get this paper in your agent:

hf papers read 2605\.17672

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.17672 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.17672 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.17672 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning

arXiv cs.CL

This paper proposes STOP (SuperTOken for Pruning), a systematic framework for pruning inefficient reasoning paths early in parallel reasoning with Large Reasoning Models. The method achieves superior efficiency and effectiveness across models from 1.5B to 20B parameters, boosting GPT-OSS-20B accuracy on AIME25 from 84% to 90% under fixed compute budgets.

Cut Your Losses! Learning to Prune Paths Early for Efficient Parallel Reasoning

Hugging Face Daily Papers

This paper introduces STOP (Super Token for Pruning), a lightweight method that learns to prune unpromising reasoning paths early during parallel decoding by appending learnable tokens and reading KV cache states, achieving 70% token reduction while improving performance on AIME and GPQA benchmarks.