PRISM Edit: One Vector for All Temporal Answers

2026-07-13

AI Summary

This work addresses the challenge that large language models struggle to update time-varying facts while preserving historically accurate answers. The study reveals, for the first time, an intrinsic two-stage mechanism within models: a time-agnostic entity representation followed by temporal modulation. Building on this insight, the authors propose an efficient editing method that requires no architectural modifications—activating the model’s inherent temporal modulation pathway via a single polysemous vector to enable temporally consistent predictions across contexts. Combining causal tracing, MLP-layer representation optimization, and a new temporal editing benchmark, TimeConflict, the method achieves a 23.3-point gain in temporal consistency on TimeConflict and a 33.7-point improvement in current-time accuracy on the temporally augmented CounterFact dataset, while running over twice as fast as the strongest baseline during inference.

Abstract

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.

Problem

• model editing
• temporal facts
• temporal consistency
• large language models
• time-aware reasoning

Innovation

• model editing
• temporal reasoning
• causal tracing
• polysemous representation
• LLM updating