EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory
EngramEdit updates facts through conditional memory while leaving the Transformer backbone fixed.
The paper says the method computes target memory representations for an updated fact across multiple phrasings, then adjusts shared n-gram embeddings to match them. It penalizes changes to embeddings that are reused often, aiming to avoid collateral changes to unrelated knowledge. In experiments, the authors report near-perfect editing success, transfer to unseen expressions, and stronger multi-hop reasoning under chain-of-thought prompting than the best baseline. Unrelated knowledge and general capabilities are described as largely preserved as edits accumulate. ArXiv · AI/CL/LG's note
The paper says the method computes target memory representations for an updated fact across multiple phrasings, then adjusts shared n-gram embeddings to match them. It penalizes changes to embeddings that are reused often, aiming to avoid collateral changes to unrelated knowledge. In experiments, the authors report near-perfect editing success, transfer to unseen expressions, and stronger multi-hop reasoning under chain-of-thought prompting than the best baseline. Unrelated knowledge and general capabilities are described as largely preserved as edits accumulate. ArXiv · AI/CL/LG's note
score 5