Knowledge Editing

Knowledge editing refers to techniques for directly modifying or inserting specific facts into a trained language model’s weights without full retraining, typically by locating and updating the parameters associated with a fact. Research in this area has found that such edits often stay locally confined — the model answers direct queries about the new fact correctly but fails to use it in downstream multi-hop reasoning, revealing that edited knowledge doesn’t automatically propagate through the model’s broader inferential chains.