Reproduce “LLM Unlearning Reveals a Stronger-Than-Expected Coreset Effect in Current Benchmarks” 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
Soumyadeep Pal, Changsheng Wang, James Diffenderfer, Bhavya Kailkhura, Sijia Liu
Research from Michigan State University, Lawrence Livermore National Laboratory, and IBM Research shows that effective LLM unlearning can be achieved with as little as 1-5% of the full forget dataset across various benchmarks and methods, often yielding comparable unlearning effectiveness and utility. The findings suggest that current unlearning benchmarks contain significant data redundancy, yet unlearned models with smaller coresets may be more susceptible to relearning during downstream fine-tuning.
Key insights
- A prominent "coreset effect" exists in LLM unlearning, where 1-5% of the forget data can achieve unlearning effectiveness comparable to using the entire dataset.
- Simple random selection of coreset data points performs comparably to more sophisticated heuristic selection methods, suggesting substantial redundancy and the driving role of salient keywords in current benchmarks.
- Achieving effective coreset unlearning often necessitates significantly more training epochs compared to unlearning with the full forget set.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.