Two new research papers explore methods to defend large language models (LLMs) against jailbreaking attacks. The first paper, "Bait-and-Recover," proposes a weight-level defense that poisons internal refusal signals to disrupt attacker edit searches, significantly increasing the refusal rate against white-box editing jailbreaks while preserving general capabilities. The second paper, "Behind Harmful Compliance," investigates different methods of inducing harmful compliance in LLMs, finding that while supervised fine-tuning, RLVR, and refusal-feature ablation can all lead to near-ceiling harmfulness, they result in distinct behavioral and mechanistic divergences. Notably, RLVR models exhibit capability-blind compliance and respond unusually to safety-reflection prompts despite direct prompt compliance. AI
IMPACT These papers introduce new techniques for enhancing LLM safety and understanding model behavior, potentially leading to more robust and secure AI systems.
RANK_REASON Two academic papers published on arXiv detailing novel methods for LLM safety and analysis.
- arXiv
- Bait-and-Recover
- CatalyzeX
- DagsHub
- Gotit.pub
- Hugging Face
- Llama-3.1:8b
- Md Rysul Kabir
- qwen2.5:7b
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
