unlearning

Unlearning removes specific knowledge or capabilities from trained models for safety, privacy, or compliance. **Motivations**: Remove copyrighted content, forget personal data (GDPR right to erasure), eliminate harmful capabilities, remove sensitive information. **Approaches**: **Fine-tuning to forget**: Train on "forget" examples with reversed labels or random outputs. **Gradient ascent**: Increase loss on data to unlearn (opposite of learning). **Representation surgery**: Edit embeddings to remove specific concepts. **Influence functions**: Approximate effect of removing specific training examples. **Challenges**: **Verification**: How to confirm knowledge is truly removed, not just suppressed? **Generalization**: Unlearn from paraphrased queries too. **Capability preservation**: Don't damage related useful capabilities. **Relearning risk**: Knowledge may resurface with prompting. **Distinction from editing**: Editing changes facts, unlearning removes them entirely. **Applications**: Copyright compliance, privacy (remove PII), safety (remove harmful knowledge). **Current state**: Active research, no foolproof methods, red-teaming needed to verify. **Tools**: Various research implementations, tofu benchmark. Important for responsible AI deployment.