ROPD: Routing-based On-Policy Distillation for LLM Safety Realignment
Source
arxiv.org
Author
Yongjian Guo, Wanlun Ma, Lingyu Shen, Xi Xiao, Sheng Wen
Date
Why it matters
Safety-realignment defenses largely assume you know the attacker's prompt template; without it they degrade badly and realigned models can be re-jailbroken by a system-prompt swap. Distribution-level realignment holds up better.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
