Research question
Do unlearning methods survive routine model compression?
Key findings
- Quantising an unlearned model to 4 bits restores about 83% of forgotten knowledge - Proposes quantisation-robust unlearning
Catastrophic Failure of LLM Unlearning via Quantization
Found that unlearning is undone by ordinary quantisation, showing that safeguard removal need not be adversarial.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.