Search papers, labs, and topics across Lattice.
This paper introduces PRISM Edit, a novel approach to model editing that addresses the limitations of the locate-and-edit paradigm by allowing large language models (LLMs) to maintain both current and historical answers for temporal facts. By leveraging causal tracing, the authors demonstrate that LLMs can utilize a two-stage internal computation to differentiate between time-agnostic representations and their temporal contexts, enabling more accurate updates. The evaluation on the newly introduced TimeConflict benchmark shows that PRISM Edit significantly outperforms existing methods, achieving a +23.3 improvement in Temporal Consistency and a +33.7 increase in Current Relative-time Score, while also being over twice as fast.
Temporal updates in LLMs can be made without sacrificing historical accuracy, achieving over 23% improvement in consistency with a single optimized representation.
Model editing keeps large language models (LLMs) up to date without retraining, but temporal facts expose a limitation of the prevailing locate-and-edit paradigm: an update is not always a replacement. When a fact changes, the new answer should become current while the old answer may remain correct in historical time contexts. Building on this insight, we use causal tracing to show that LLMs already support this distinction via a two-stage internal computation: early MLP layers retrieve a time-agnostic subject representation, and later layers modulate it with temporal context to yield the time-correct answer. Motivated by this finding, we introduce PRISM Edit, which optimizes a single polysemous representation across temporal contexts and leverages the model's inherent modulation pathway to route it to temporally correct predictions, without any architectural modification. We evaluate on TimeConflict, a new temporal editing benchmark we introduce, and on temporally augmented CounterFact. PRISM Edit improves over the best baseline by +23.3 Temporal Consistency (TC) and +33.7 Current Relative-time Score (CRS) on average while being more than 2x faster. Code and data are publicly available at https://github.com/AnonymousStudy972/PRISM-Edit.
Tsinghua AI