PRISM Edit: One Vector for All Temporal Answers

PRISM Edit addresses a central weakness of model editing for changing facts: a new answer should be current, while the old answer may remain correct in historical contexts. Using causal tracing, the authors report a two-stage mechanism in which early MLP layers retrieve a time-agnostic subject representation and later layers modulate it with temporal context. PRISM Edit optimizes one shared, polysemous representation and relies on this existing pathway without architectural changes. On the new TimeConflict benchmark and temporally augmented CounterFact, the paper reports average gains of 23.3 Temporal Consistency points and 33.7 Current Relative-time Score points over the strongest baseline, while running more than twice as fast.