Reproduce “Forget Narrowly, Retain Broadly: Unlearning as an Asymmetric Generalization Problem” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Amit Peleg, Naman Deep Singh, Naama Pearl, Bibhabasu Mohapatra, Matthias Hein
Peleg et al. introduce SUITE, a new evaluation protocol and training corpus for LLM unlearning designed to rigorously test for under-forgetting and over-forgetting, and present JensUn++, an unlearning algorithm that achieves a superior forget-retain trade-off by generating natural refusal responses.
Key insights
- Unlearning in LLMs constitutes an "asymmetric generalization problem" requiring both intensive generalization for forgetting and extensive generalization for retention.
- The quality and structure of unlearning training data (like SUITE) are as crucial as algorithmic design for achieving effective and reliable unlearning.
- Traditional unlearning evaluations often overestimate forgetting and mask over-forgetting, failing to detect knowledge leakage via diverse queries or collateral damage to related information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.