PRISM Edit: One Vector for All Temporal Answers

Traditional knowledge editing methods treat facts as static, often overwriting historical information when updating a model with new temporal data. This leads to "catastrophic forgetting" of past facts. The authors investigate whether large language models (LLMs) possess an internal mechanism to distinguish between time-agnostic subjects and time-specific contexts, and whether this mechanism can be leveraged for more effective temporal editing.

Using causal tracing, the authors map how LLMs process temporal knowledge. They discover a two-stage computation: early MLP layers retrieve a time-agnostic subject representation, while later attention layers modulate this signal with temporal context to produce the final output. Based on this, they introduce PRISM Edit, which avoids the traditional "locate-and-edit" trap of overwriting facts with a single, time-blind value. Instead, it optimizes a single "polysemous" vector that encodes multiple temporal answers, allowing the model's own downstream modulation to route the representation to the correct answer based on the provided time context.

PRISM Edit significantly outperforms existing state-of-the-art editing methods on the newly introduced TIMECONFLICT benchmark and temporally augmented COUNTERFACT. It achieves a +23.3 improvement in Temporal Consistency (TC) and a +33.7 improvement in Current Relative-time Score (CRS) compared to the best baseline. Because it does not require architectural modifications or auxiliary modules, it is also more than twice as fast as existing temporal editing strategies.

This research demonstrates that effective temporal editing does not require adding new parameters or complex external structures. By aligning editing objectives with the model's existing internal information flow, researchers can update LLMs to reflect changing world knowledge while preserving historical accuracy, a critical requirement for deploying models in dynamic, real-world environments.