Scholarlink

<p>With the rapid development of large language models (LLMs), model safety has been a critical concern in key domains such as the humanities, math, and education. Existing safety alignment methods that rely on supervised fine-tuning or preference optimization are generally effective against single-turn jailbreak attacks. However, they remain vulnerable to multi-turn jailbreaks that accumulate inductive risk across dialogue turns and progressively erode safety constraints. To address this limitation, we propose Progressive Induction-Aware Representation Separation (<b>PIARS</b>), a novel optimization framework that explicitly quantifies and mitigates inductive risks under multi-turn jailbreaks. <b>PIARS</b> first introduces an induction degree that measures how risk accumulates across turns. This degree is integrated into a multi-turn safety objective that enforces representation separation between safe and unsafe dialogue states, improving discriminative ability against contextual drifts. To ensure that this separation remains both effective and practical, <b>PIARS</b> augments the objective with auxiliary constraints for safety boundaries and utility preservation. Notably, we progressively adjust the multi-objective weights to achieve balance between safety and utility during the optimization process. Extensive experiments on several widely used LLMs demonstrate that <b>PIARS</b> achieves a superior defense against multi-turn jailbreaks, reducing attack success rates by over 50% on average compared to strong baselines, while simultaneously lowering over-refusal and preserving model utility. Our code is available at: <a href="https://github.com/jsj2024/PIARS">https://github.com/jsj2024/PIARS</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Progressive Induction-Aware Optimization for LLMs Safety Under Multi-turn Jailbreaks

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Sijia Jing,
  • Hongye Cao,
  • Boyan Wang,
  • Jing Huo,
  • Yang Gao

摘要

With the rapid development of large language models (LLMs), model safety has been a critical concern in key domains such as the humanities, math, and education. Existing safety alignment methods that rely on supervised fine-tuning or preference optimization are generally effective against single-turn jailbreak attacks. However, they remain vulnerable to multi-turn jailbreaks that accumulate inductive risk across dialogue turns and progressively erode safety constraints. To address this limitation, we propose Progressive Induction-Aware Representation Separation (PIARS), a novel optimization framework that explicitly quantifies and mitigates inductive risks under multi-turn jailbreaks. PIARS first introduces an induction degree that measures how risk accumulates across turns. This degree is integrated into a multi-turn safety objective that enforces representation separation between safe and unsafe dialogue states, improving discriminative ability against contextual drifts. To ensure that this separation remains both effective and practical, PIARS augments the objective with auxiliary constraints for safety boundaries and utility preservation. Notably, we progressively adjust the multi-objective weights to achieve balance between safety and utility during the optimization process. Extensive experiments on several widely used LLMs demonstrate that PIARS achieves a superior defense against multi-turn jailbreaks, reducing attack success rates by over 50% on average compared to strong baselines, while simultaneously lowering over-refusal and preserving model utility. Our code is available at: https://github.com/jsj2024/PIARS.