arXiv：cs.AI· Haoming Yang, Ke Ma, Yangbangyan Jiang, Ligong Zhang, Xiaojun Jia, Longtao Huang, Yingfei Sun, Qianqian Xu, Qingming Huang·· 16 小时前
LLM 推理中的叙事悖论：PDRA 攻击如何将推理轨迹引向失准
Narrative Paradoxes in LLM Inference: Shaping Reasoning Trajectories toward Misalignment
AI 导读
研究者提出 PDRA（Paradox-driven Reasoning-based Attack），通过将恶意行为者与良性任务配对构建悖论式叙事核心，再嵌入强制三阶段分析框架（表层目的、潜在意图、操作重建），从而将 LLM 的多步推理轨迹编程为有害路径。该攻击使模型内部表征偏离拒答相关模式，抑制安全触发词汇并激活规划类词汇，在多个 LLM 上取得 SOTA 攻击成功率且效率较高。
正文
Abstract:Explicit reasoning in large language models (LLMs) is often associated with reliability, yet it can still pose risks when a model infers harmful goals under a benign context. We identify a distinct challenge: ensuring the safety of the \emph{multi-step reasoning process} itself, where locally coherent steps may still lead to globally harmful outcomes. To explore this vulnerability, we propose PDRA (Paradox-driven Reasoning-based Attack), a novel attack method that programs a harmful reasoning trajectory. PDRA first constructs a paradoxical narrative core by pairing a malicious actor (e.g., a terrorist) with a benign task (e.g., writing a safety bulletin). It then embeds this core into a structured analytical framework that forces the model through a mandatory three-stage analysis -- surface purpose, latent intent, and operational reconstruction -- guiding it to produce detailed harmful content while maintaining narrative coherence. Mechanistically, we show PDRA steers the model's internal representations away from refusal-related patterns, suppressing safety-triggering lexicons while activating planning-oriented terms. Empirically, PDRA achieves state-of-the-art attack success rates across diverse LLMs with high efficiency. \textcolor{red}{\textbf{Warning:} This paper may contain sensitive content.}
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.16312 [cs.AI] |
| (or arXiv:2610.16312v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.16312 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Haoming Yang [view email]
[v1]
Mon, 12 Oct 2026 07:38:37 UTC (1,670 KB)
来源：arXiv：cs.AI · arxiv.org