Measure, Don't Optimize: Forecasting Recovery in LLM Unlearning

Shared on Bluesky by 2 AI experts

• A arxiv cs.CL @arxiv-cs-cl.bsky.social: Zirui Song, Huaxing Liu, Xiang Wang, Shuai Li, Xinye Li, Lang Gao, Jinghui Zhang, Zheng Lu, Fengxian Ji, Xiaojun Chang, Xiuying Chen Measure… →
• A AI Firehose @ai-firehose.column.social: This study presents J-Access, a method to audit residual knowledge accessibility in unlearned language models. It predicts recovery risk eff… →