knowledge editing

Knowledge editing updates a model's stored factual knowledge without expensive full retraining. **Why needed**: Facts change (new president, updated statistics), training data had errors, personalization requirements. **Knowledge storage hypothesis**: MLPs in middle-late layers store key-value factual associations. Editing targets these parameters. **Methods**: **ROME (Rank-One Model Editing)**: Identify layer storing fact, compute rank-one update to change association. **MEMIT**: Extends ROME to batch edit thousands of facts. **MEND**: Meta-learned editor network. **Locate-then-edit**: First find responsible neurons, then update. **Edit specification**: State change as (subject, relation, old_object → new_object). Model should answer queries about subject with new object. **Challenges**: **Generalization**: Handle paraphrases of the query. **Locality**: Don't break other knowledge. **Coherence**: Related knowledge stays consistent. **Scalability**: Many edits accumulate issues. **Evaluation benchmarks**: CounterFact, zsRE. **Comparison to RAG**: RAG keeps knowledge external (easier updates), editing modifies model (no retrieval latency). **Limitation**: Only works for factual knowledge, not complex reasoning or skills.