arxiv.org web signal

CleanSlate: picking what to forget breaks LLM unlearning

TL;DR

  • A new benchmark called CleanSlate tests whether an LLM can be made to stop reproducing specific songs or books on request.
  • Simple lexical or exact-substring methods for choosing what to forget produce forget sets that lead to weak suppression.
  • An evaluation-aware curator almost fully suppresses the target output but damages unrelated content and general model capabilities.

Machine unlearning research usually starts with a list of examples the model should forget. A new paper argues that list is the real problem.

"In realistic language-model deployments, a requester may ask a model to stop reproducing a song or book without knowing which spans, documents, quotations, or near-duplicates in a trillion-token corpus support that behavior," write Animesh Jha, Arpandeep Khatua, Youssef Allouah, and Sanmi Koyejo in a paper posted to arxiv. They call the upstream step "forget set curation" and introduce a benchmark, CleanSlate, for verbatim output suppression over songs and books, with "model-specific extraction profiles, content-grounded QA, and capability-retention evaluations."

Neither of the curators they test comes out well. Simple approaches, they report, "often yield forget sets that lead to weak suppression." A smarter one "suppresses requested continuations almost completely, but causes collateral regression on non-requested content and model-dependent capability loss."

The authors' bottom line: "practical unlearning is not only an optimization problem once a forget set is given: the data chosen for forgetting determines both what can be unlearnt and what else is damaged."

The abstract names no models tested and publishes no per-song suppression numbers. Two researchers we follow shared it within a day.

Shared on Bluesky by 2 AI experts