New SEMA framework enhances multi-turn jailbreak attacks on LLMs

Researchers have developed SEMA, a novel framework designed to improve multi-turn jailbreak attacks against large language models. SEMA utilizes a two-stage process: prefilling self-tuning to generate structured adversarial prompts and reinforcement learning with an intent-drift-aware reward to maintain harmful objectives. This approach achieves state-of-the-art attack success rates, outperforming existing methods by a significant margin on benchmarks like AdvBench. AI

IMPACT This research provides a more robust method for red-teaming LLMs, potentially accelerating the identification and mitigation of safety vulnerabilities.

RANK_REASON The cluster contains an academic paper detailing a new method for LLM safety research. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• AdvBench
• Direct Preference Optimization
• Mingqian Feng
• supervised fine-tuning

• safety
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →