当推理陷入迷途：失控推理的注意力动态演化与实时干预

When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

文章背景与核心概要

随着大推理模型 (Large Reasoning Models, LRMs) 的兴起，依靠延长思维链进行深度思考已成为攻克高难度数学与逻辑难题的核心范式，但这种漫长的思维过程极易误入歧途，蜕变为无休止的冗余验证与无限生成死循环，导致推理开销激增与算力资源枯竭。针对现有多依赖生硬截断长文本或浅层重复检测等被动防御手段的局限，研究团队提出了创新的 RADAR (Reasoning-state Analysis via Dynamic Attention Responses) 分析框架。该框架首次将大模型的推理生成过程形式化解构为四大离散状态，不仅能在推理运行时实时判定思维状态，更从底层注意力机制上清晰刻画了良性反思如何一步步恶化为失控生成。在此基础上，研究进一步提出了无需二次微调训练的 Attention Realignment 实时干预机制，在表层文本出现显性重复之前便将异常注意力分布导向回正常基准，在几乎无损模型原生解题能力的前提下成功消除了死循环，为保障大推理模型的高效稳定运行开辟了全新路径。

概述与核心摘要

Overview & Summary

本文深入研究了 Large Reasoning Models (LRMs) 的失效模式。这类大推理模型往往借助更长推理步骤的“深度思考”在复杂任务中斩获优异表现，但这种漫长的思维延伸过程经常会退化为冗余的重复验证循环与无限生成状态。这些失控的推理行为不仅会成倍推高推理成本，带来计算资源枯竭的风险，还会严重拉低线上服务的响应质量。

This paper investigates the failure modes of Large Reasoning Models (LRMs), which often leverage extended reasoning to excel at complex tasks. However, this prolonged thinking process can frequently degenerate into redundant verification loops and persistent generation cycles. These uncontrolled behaviors drive up inference costs, risk resource exhaustion, and degrade service quality.

为了打破现有被动应对策略 (例如直接暴力截断输出或仅针对字面词句做浅层重复检测) 的局限，作者提出了 RADAR (Reasoning-state Analysis via Dynamic Attention Responses) 框架。这一创新框架带来了以下关键突破： * 四状态形式化解构：将 LRM 的生成过程细分为四种界限分明的思维状态； * 实时状态判定：在模型推理过程中实时识别当前所处的推理状态； * 机理解析与演化映射：从注意力机理上剖析有益的自我反思是如何逐步蜕化为失控生成的； * 引入 Attention Realignment：提出了一种精准的纠偏干预策略，将偏离常态的异常注意力分布拉回健康基准模式——在几乎不损耗模型任务求解能力的前提下，成功阻断陷入死循环。

To address the limitations of existing reactive mitigations (such as output truncation or superficial repetition detection), the authors introduce RADAR (Reasoning-state Analysis via Dynamic Attention Responses). This novel framework: * Operationalizes LRM generation into four distinct states. * Identifies current reasoning states in real-time. * Mechanistically maps how healthy reflection devolves into uncontrolled generation. * Introduces Attention Realignment, a corrective intervention that shifts abnormal attention distributions back toward healthy patterns—successfully reducing looping while preserving task performance.

论文元数据与发表详情

Metadata & Publication Details

• arXiv 编号： arXiv:2609.38817 [cs.AI]
• 研究领域： 人工智能 (cs.AI) ；计算与语言 (cs.CL)
• 提交时间： 2026 年 9 月 30 日
• 数字对象唯一标识符 (DOI) ： 10.48550/arXiv.2609.38817
• 知识共享许可： Creative Commons Attribution 4.0

• arXiv ID: arXiv:2609.38817 [cs.AI]
• Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
• Submitted on: September 30, 2026
• DOI: 10.48550/arXiv.2609.38817
• License: Creative Commons Attribution 4.0

论文作者

Authors

• Yuanhe Zhang
• Ziwei Wang
• Jie Ren
• Haoran Gao
• Zhenhong Zhou
• Fanyu Meng
• Cong Wu
• Li Sun
• Sen Su

• Yuanhe Zhang
• Ziwei Wang
• Jie Ren
• Haoran Gao
• Zhenhong Zhou
• Fanyu Meng
• Cong Wu
• Li Sun
• Sen Su

论文摘要

Abstract

大推理模型 (Large Reasoning Models, LRMs) 通过展开长链条的扩展推理，显著提升了攻克高难度复杂任务的能力；然而，这一思维延伸过程也极易“走入死胡同”，蜕变成无休止的冗余验证与无限生成循环。这种失控推理现象不仅成倍推高了推理成本，更伴随着算力资源枯竭与服务质量降级的巨大风险。然而，现有的应对手段大多非常被动，要么简单粗暴地直接截断超长输出，要么仅能迟钝地检测字面上的浅层重复，既无法甄别模型究竟是在进行正常的深度探索还是已经深陷思维泥潭，更无法解释良性推理究竟是如何蜕变为有害行为的。在本文中，我们将 LRM 的生成过程形式化解构为四种离散状态，并提出了基于动态注意力响应的推理状态分析框架 —— RADAR (Reasoning-state Analysis via Dynamic Attention Responses) ，该框架不仅能够实时识别模型当前的推理状态，还能从底层机理上描绘出有效反思演变为失控生成的过程。在 RADAR 分析规律的引导下，我们进一步探索了 Attention Realignment 干预机制，将偏离常态的异常注意力分布重新对齐拉回至正常请求的基准分布，并系统检验了这种纠正手段对抑制过度反思和持续循环的有效性。时序分析表明，失控推理的核心特征在于其注意力分布与正常生成轨迹发生明显偏离，且这种异常趋势在字面重复开始之前便已显现并可被精准检测。通过 Attention Realignment 修正这些注意力偏离，能够在近乎完全保留模型固有优异性能的同时，稳定、显著地消除循环现象。总体而言，RADAR 从底层机理层面揭示了模型推理陷入失控的物理演进过程，为精准识别关键失效阶段并设计高针对性的运行时动态干预机制提供了极具实操价值的明确指引。

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

核心贡献与创新点

Key Contributions

1. 四状态形式化解构 (Four-State Operationalization) ： 将 LRM 文本生成过程明确细分为四种不同状态，清晰隔开了有益的建设性自我反思与病态的死循环生成；
2. RADAR 诊断框架 (RADAR Framework) ： 基于动态注意力响应特征，实现对模型推理状态的高精度实时识别与监测；
3. 早期异常预警 (Early Detection) ： 实验证实注意力分布的异常偏离在表层文本出现明显重复之前便已发生，能够实现超前预警与提前拦截；
4. Attention Realignment 纠偏干预： 提出了一种注意力重对齐干预机制，能够将偏离常态的异常注意力分布重新导向回基准模式，在几乎不折损模型原有解题能力的同时，高效遏制失控生成死循环。

1. Four-State Operationalization: Categorizes LRM text generation into distinct states to separate constructive reflection from pathological looping.
2. RADAR Framework: Provides real-time identification of reasoning states using dynamic attention responses.
3. Early Detection: Demonstrates that attention anomalies occur and can be detected before overt surface-level repetition begins.
4. Attention Realignment: A corrective intervention that steers abnormal attention distributions back to baseline norms, effectively mitigating runaway generation loops while retaining model capability.