CoFiBreak：面向多轮越狱的由粗到细多样性奖励
原标题：CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn Jailbreak
AI 摘要
论文提出 CoFiBreak，一种用于 GRPO 的由粗到细多样性奖励方法，通过粗粒度奖励鼓励攻击者使用组内稀缺策略、细粒度奖励在去除有害意图方向后惩罚模板复用，从而在保持攻击成功率的同时提升多轮越狱攻击的多样性。以 Llama3.1-8B-Instruct 为受害模型训练后，在 HarmBench 上发现的细粒度成功策略数量为最强基线的 1.6–2.0 倍，在 AdvBench 上达到 98.8% ASR@10 且多样性得分为最强攻击基线的 2.1 倍，并可迁移至 Claude-Opus-4.6 达到 71.0% ASR@10。
正文节选
CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn Jailbreak Abstract Automated red teaming trains an attacker large language model (LLM) to elicit harmful responses from a victim LLM, but reinforcement learning can collapse attackers onto a narrow set of prompts when attack success is the only reward. Such a collapse of attack diversity can bias safety evaluation toward repeatedly testing near-duplicate attack patterns rather than the underlying harmful intents. Existing met