RepSelect aims to make LLM unlearning harder to reverse

A new arXiv paper proposes RepSelect, a method for more robust LLM unlearning. The authors argue that existing techniques often forget shallowly and can be reversed through fine-tuning or few-shot prompting.

RepSelect tries to isolate representations specific to the forget set while leaving general capabilities intact. That addresses a key tradeoff in unlearning: removing targeted knowledge without damaging the rest of the model.

The work is relevant for compliance, safety, and model governance, where claims that a model has forgotten something need to withstand adversarial recovery attempts.