New CoFiBreak Method Targets Multi-Turn Jailbreaks in LLM Red Teaming
Researchers introduce CoFiBreak, a coarse-to-fine diversity reward designed to prevent reinforcement learning collapse in automated red teaming, enabling attacker LLMs to more effectively elicit harmful responses from victim LLMs across multiple turns.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-02 05:00 · arXiv cs.CL
CoFiBreak: A Coarse-to-Fine Diversity Reward for Effective Multi-Turn Jailbreak