K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models
Two consistent dissociations between representation-level alignment and behavioral expression are reported, plus a common failure under position perturbations, characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.
Yu Sun, Meng Lu, Cong Feng et al.
· 0 citations