Seminars

Home Schools Computational Sciences Seminars

TITLE: What Does It Mean for an LLM to Forget? Recent Progress in Machine Unlearning

ABSTRACT: Machine unlearning asks how to remove the influence of a precisely specified forget set from a trained model, ideally yielding behavior indistinguishable from retraining without that data. I will begin with the classical “data deletion” formulation from the classification literature and discuss why strong (retrain-equivalent) unlearning is difficult to evaluate in deep networks, motivating our recent information-theoretic evaluation metric. I will then review recent developments in LLM unlearning, focusing on three practical evaluation gaps highlighted by my recent work: i) unlearning in large reasoning models with step-wise trace evaluation (R-TOFU), ii) mixed prompts where forget/retain requests co-occur (SEPS), and iii) realistic overlap where unlearning must remove unique content while preserving shared facts (DUSK). Finally, I will argue that “machine unlearning” is often used as an umbrella for policy-driven suppression/refusal/editing, and I will outline clearer terminology and reference-based evaluation criteria aligned with the intended guarantee.