Off-Policy Replay method improves LLM unlearning efficiency

Arxiv CS.CL (NLP/language models)

Proposes Off-Policy Replay, an RL-based method for efficient LLM unlearning that removes hazardous knowledge without full retraining. Demonstrates cost-effective alternative while preserving general utility in pretrained models.

LLM

Arxiv CS.CL (NLP/language models)·Zirui PangChenlong ZhangHaosheng TanZhuoran JinJiaheng WeiZixin Zhong

22 hours ago