EngramEdit
A knowledge editing method for large language models based on a conditional memory architecture. It decouples factual knowledge storage from general model computation to enable precise parametric knowledge updates while protecting unrelated knowledge from unintended modification.
Timeline (last 90 days)
EngramEdit is a method proposed in a paper that achieves decoupled updates of factual knowledge in large models through conditional memory, strictly limiting the scope of changes to the knowledge storage module while keeping the Transformer backbone completely frozen.
EngramEdit's method consists of two steps: the first step computes the target memory representation so that the model can predict the new fact across multiple expressions of that fact, and the second step performs joint updates on the shared n-gram embeddings.
EngramEdit applies stronger update penalties to embeddings that are frequently reused in order to maximally preserve the integrity of unrelated knowledge.
EngramEdit achieves an editing success rate approaching 100%.
Under chain-of-thought (CoT) prompting, EngramEdit's multi-hop reasoning accuracy approaches three times that of the strongest baseline method.
The revised knowledge takes effect not only in sentence patterns seen during training but also in expressions never encountered before.
EngramEdit repositions the role of the conditional memory architecture from a mere model scaling tool to an editable knowledge interface.
All Facts (7)
EngramEdit is a method proposed in a paper that achieves decoupled updates of factual knowledge in large models through conditional memory, strictly limiting the scope of changes to the knowledge storage module while keeping the Transformer backbone completely frozen.
50%UnverifiedEngramEdit's method consists of two steps: the first step computes the target memory representation so that the model can predict the new fact across multiple expressions of that fact, and the second step performs joint updates on the shared n-gram embeddings.
50%UnverifiedEngramEdit applies stronger update penalties to embeddings that are frequently reused in order to maximally preserve the integrity of unrelated knowledge.
50%UnverifiedEngramEdit achieves an editing success rate approaching 100%.
50%UnverifiedUnder chain-of-thought (CoT) prompting, EngramEdit's multi-hop reasoning accuracy approaches three times that of the strongest baseline method.
50%UnverifiedThe revised knowledge takes effect not only in sentence patterns seen during training but also in expressions never encountered before.
50%UnverifiedEngramEdit repositions the role of the conditional memory architecture from a mere model scaling tool to an editable knowledge interface.
50%