arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.02600cs.LGcs.AI
收藏

自回归与扩散语言模型中的逐步拒绝动态

Step-Wise Refusal Dynamics in Autoregressive and Diffusion Language Models

Department of Computer Science, Technion – Israel Institute of Technology(技术学院计算机科学系,以色列技术学院);INSIGHT Lab, School of Electrical and Computer Engineering, Ben-Gurion University of the Negev, Israel(内斯坦实验室,贝内-加隆大学内加尔分校,以色列);Computer Science Department, University of Haifa, Haifa, Israel(海法大学计算机科学系,海法,以色列)

Eliron Rahimi, Elad Hirshel, Rom Himelstein, Amit LeVi, Avi Mendelson, Chaim Baskin

AI总结:

研究扩散语言模型(DLM)与自回归(AR)模型在拒绝有害生成行为上的差异,发现扩散重掩码机制可促进恢复,提出逐步拒绝内部动态(SRI)信号,并基于此构建无需修改推理的越狱检测器。

摘要:

扩散语言模型(DLM)最近已成为自回归(AR)模型的有竞争力的替代方案,提供并行解码、竞争性生成质量以及越狱鲁棒性改善的初步证据。尽管取得了这些进展,但采样机制在塑造拒绝行为中的作用仍知之甚少。为填补这一空白,我们提出了一项关于逐步拒绝动态的全面研究。我们表明,扩散重掩码可以促进从有害中间生成中恢复,提供证据表明这种行为与采样机制相关,并证明从AR采样切换到扩散采样可提高越狱鲁棒性,包括在固定模型权重下。为了捕捉在文本层面不可观察的生成动态,我们提出了逐步拒绝内部动态(SRI)信号。与我们的文本层面发现一致,SRI表明恢复主要在AR采样下失败,这些失败在SRI空间中通常相对于无害生成表现为异常。基于这一观察,我们表明SRI能够实现一个简单的越狱检测器,该检测器无需修改推理,并且仅通过在良性SRI信号上训练即可泛化到未见攻击。我们的评估表明,该检测器匹配或超越现有越狱检测基线,同时增加可忽略的开销。

英文摘要:

Diffusion language models (DLMs) have recently emerged as a competitive alternative to autoregressive (AR) models, offering parallel decoding, competitive generation quality, and initial evidence of improved jailbreak robustness. Despite this progress, the role of sampling mechanisms in shaping refusal behavior remains poorly understood. To address this gap, we present a comprehensive study of step-wise refusal dynamics. We show that diffusion remasking can promote recovery from harmful intermediate generations, provide evidence that this behavior is tied to the sampling mechanism, and demonstrate that switching from AR to diffusion sampling improves jailbreak robustness, including under fixed model weights. To capture generation dynamics not observable at the text level, we propose the Step-Wise Refusal Internal Dynamics (SRI) signal. Consistent with our text-level findings, SRI shows that recovery fails primarily under AR sampling, with these failures often appearing anomalous relative to harmless generations in the SRI space. Based on this observation, we show that SRI enables a simple jailbreak detector that does not modify inference and generalizes to unseen attacks by training only on benign SRI signals. Our evaluation shows that this detector matches or outperforms existing jailbreak detection baselines while adding negligible overhead.

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。

↑