PRISM Edit unlocks temporal LLM editing

While most model editors treat memory updates as a zero-sum game, PRISM Edit's utilization of polysemous vectors acknowledges that time-bound facts coexist rather than overwrite one another.

* By optimizing a single vector across different temporal contexts, PRISM Edit aligns with LLMs' innate two-stage mechanisms (retrieval then temporal modulation).

* The technique avoids modifying the model's architecture, making it highly portable and compatible with standard LLM weights.

* Demonstrating superior performance on the TimeConflict and CounterFact benchmarks, it is over 2× faster than existing locate-and-edit baselines.

* However, scaling this to long histories or complex, overlapping temporal events remains an open challenge for vector-based memory modulation.