Explain the superiority of partial distractor suppression

Determine why removing half of each distractor’s letter-identity content improves sensitivity more than completely removing that content in the linear interference model of transformer readout.

Background

The paper models the decision variable as a signal for whether the query matches the target plus interfering letter content contributed by non-target items. Empirically, suppressing half of each distractor’s letter-identity component improves sensitivity more than sharpening attention and reduces false alarms while largely preserving hit rates.

The authors note that the theoretical proposition predicts the benefits of distractor suppression but does not account for the observed non-monotonic intervention effect: half-strength removal is more beneficial than complete removal. Thus, the mechanism explaining why partial suppression is optimal remains unresolved within the paper’s linear readout model.