SafeProbing: Probing Hidden Safety Signals During Decoding to Defend Against LLM Jailbreaks
[Daily Paper] In-Decoding Safety-Awareness Probing: Surfacing Hidden Safety Signals to Defend LLMs Against Jailbreaks
AI summary
SafeProbing probes hidden safety signals during the decoding stage to intercept harmful generations under jailbreak. On Qwen models, it achieves an average defense success rate of 95.1% against 7 attack methods including AutoDAN and PAIR, 15 percentage points higher than the next-best baseline at 80.1%, while probing only 20% of decoding steps randomly sampled. On XSTest, over-refusal occurs only 18 times, with zero degradation in math problem-solving performance.