arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.07470cs.AI
收藏

推理大语言模型对其推理链的干预是否具有鲁棒性?

Are Reasoning LLMs Robust to Interventions on Their Chain-of-Thought?

Helmholtz Munich(海德堡-慕尼黑研究所);Ludwig Maximilian University of Munich(慕尼黑路德维希-马克西米利安大学);Technical University of Munich, Munich Center for Machine Learning (MCML)(慕尼黑技术大学,慕尼黑机器学习中心)

Alexander von Recum, Leander Girrbach, Zeynep Akata

AI总结:

研究评估了推理大语言模型在面对不同干预时的鲁棒性,发现模型在干扰下能恢复但存在效率与准确性的权衡。

摘要:

推理大语言模型(RLLMs)在给出答案前会生成逐步的推理链(CoTs),这提高了其在复杂任务上的性能,并使推理过程更加透明。但这些推理轨迹在受到其中发生的干扰时有多鲁棒呢?为了解决这个问题,我们引入了一个受控评估框架,在固定的时间步长上扰动模型自身的CoT。我们设计了七种干预(无害、中性、对抗性)并将其应用于多个开放权重的RLLMs,涵盖数学、科学和逻辑任务。我们的结果表明,RLLMs总体上是鲁棒的,能够从各种扰动中可靠地恢复,鲁棒性随着模型大小的增加而提高,当干扰发生在早期时则会降低。然而,鲁棒性并不具有风格不变性:改写会抑制怀疑性表达并降低性能,而其他干扰会引发怀疑并支持恢复。恢复也伴随着成本:中性和对抗性噪声可以使CoT长度增加超过200%,而改写会缩短轨迹但损害准确性。这些发现提供了关于RLLMs如何维持推理完整性的新证据,识别怀疑作为核心恢复机制,并突显了鲁棒性和效率之间权衡,这将是未来训练方法需要解决的问题。

英文摘要:

Reasoning LLMs (RLLMs) generate step-by-step chains of thought (CoTs) before giving an answer, which improves performance on complex tasks and makes reasoning more transparent. But how robust are these reasoning traces to disruptions that occur within them? To address this question, we introduce a controlled evaluation framework that perturbs a model's own CoT at fixed timesteps. We design seven interventions (benign, neutral, and adversarial) and apply them to multiple open-weight RLLMs across Math, Science, and Logic tasks. Our results show that RLLMs are generally robust, reliably recovering from diverse perturbations, with robustness improving with model size and degrading when interventions occur early. However, robustness is not style-invariant: paraphrasing suppresses doubt-like expressions and reduces performance, while other interventions trigger doubt and support recovery. Recovery also carries a cost: neutral and adversarial noise can inflate CoT length by more than 200%, whereas paraphrasing shortens traces but harms accuracy. These findings provide new evidence on how RLLMs maintain reasoning integrity, identify doubt as a central recovery mechanism, and highlight trade-offs between robustness and efficiency that future training methods should address.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑