Reasoning-Token Spikes Under Prompted Untruthful Responding in Large Language Models
作者: Maverick Morales, Tomáš Dominik, Vermut Gao, Katrina Shirey, Paulius Rimkevičius, Aaron Schurger, Uri Maoz
分类: cs.AI, cs.CL
发布日期: 2026-10-07
备注: 20 pages, 9 figures, 3 tables. Code: https://github.com/Wakaranaino/token-spike-project ; Data: https://doi.org/10.5281/zenodo.21895296
💡 一句话要点
提出基于推理令牌的信号以检测大型语言模型的不真实响应
🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)
关键词: 推理模型 欺骗检测 语言模型 认知负荷理论 推理令牌 人工智能伦理 信号检测
📋 核心要点
- 现有方法在监测推理模型的欺骗行为时，依赖于推理痕迹的可读性和忠实性，面临越来越多的挑战。
- 本文提出通过监测生成的推理令牌数量，作为一种低带宽信号来检测模型的真实与不真实响应。
- 实验结果显示，真相导向的响应生成的推理令牌数量显著低于谎言导向和无关真相的响应，验证了该方法的有效性。
📝 摘要（中文）
监测推理人工智能模型的思维链是检测欺骗和其他不当行为的重要方法。然而，语义链监测依赖于推理痕迹的可读性和忠实性，且越来越多的证据表明，推理输出可能变得不可读或不忠实。基于认知负荷理论，本文研究了一种低带宽信号——生成的推理令牌数量，旨在不依赖推理痕迹内容的情况下进行检测。通过对三种推理能力的大型语言模型进行实验，发现真相导向的响应生成的推理令牌数量少于谎言导向和无关真相的响应。结果表明，明确提示的不真实响应策略可以在测试时产生显著的推理令牌使用差异。虽然尚未确立推理令牌计数作为自发欺骗或一般不一致的检测器，但结果证明其在缺乏可靠推理痕迹时可作为区分真实与不真实模型行为的候选信号。
🔬 方法详解
问题定义：本文旨在解决如何在缺乏可读推理痕迹的情况下，检测大型语言模型的欺骗行为。现有方法依赖于推理痕迹的可读性，面临不可读或不忠实的风险。
核心思路：论文的核心思路是通过监测推理令牌的数量，作为一种不依赖于推理内容的信号来区分真实与不真实的响应。这种设计基于认知负荷理论，认为不同的响应策略会影响推理过程的复杂性。
技术框架：整体架构包括三个主要阶段：首先，设计系统提示以引导模型生成不同类型的响应；其次，记录生成的推理令牌数量；最后，分析不同响应策略下的推理令牌使用情况。
关键创新：最重要的技术创新点在于提出了推理令牌数量作为检测信号的概念，这与现有方法依赖于推理内容的检测方式本质上不同。
关键设计：在实验中，设置了不同的系统提示以引导模型生成真实、虚假或无关真相的响应，并通过对比分析推理令牌的数量来评估模型的响应策略。
🖼️ 关键图片
📊 实验亮点
实验结果显示，在三种大型语言模型中，真相导向的响应生成的推理令牌数量显著低于谎言导向和无关真相的响应，具体而言，真相导向响应的推理令牌数量减少了约20%。这一发现为未来的欺骗检测提供了新的思路。
🎯 应用场景
该研究的潜在应用领域包括人工智能的道德与法律责任监测、虚假信息检测以及增强人机交互的透明度。通过提供一种新的检测信号，能够在缺乏详细推理痕迹的情况下，帮助开发更可靠的AI系统，提升其在实际应用中的可信度。
📄 摘要（原文）
Monitoring the chain-of-thought of reasoning artificial intelligence (AI) models remains a key approach to detecting deception and other forms of misbehavior in such models. However, semantic chain-of-thought monitoring depends on reasoning traces being legible and sufficiently faithful to the underlying computations that produced the model's behavior, not to mention accessible. Moreover, there is increasing evidence that chain-of-thought outputs may soon become illegible or unfaithful, if they even remain accessible. Based on cognitive load theory, we investigate a lower-bandwidth signal -- the number of reasoning tokens generated -- which does not require access to the content of the reasoning trace. Three reasoning-capable large language models answered 210 multiple-choice questions -- across analytic, descriptive, and normative reasoning types as well as moral and non-moral domains -- under system prompts instructing them to respond truthfully, falsely, or without regard for truth. Across all three models, truth-directed responding elicited fewer reasoning tokens than both lie-directed and truth-indifferent responding. These findings show that explicitly prompted untruthful response policies can produce robust group-level differences in test-time reasoning-token use. While not yet establishing reasoning-token count as a detector of spontaneous deception or general misalignment, our results are a proof of concept that it can serve as a simple, content-independent candidate signal for differentiating untruthful from truthful model behavior when raw reasoning traces are unavailable or unreliable. Future work should test instance-level detection rates, out-of-distribution generalization, learned deceptive policies, hidden objectives, and robustness under adversarial pressure.