Read-Best Is Not Steer-Best: A Probing--Steering Layer Dissociation in Omni-Modal Large Language Models
Results show that probing accuracy is a poor heuristic for selecting intervention layers and suggest a cross-architecture mid-to-late selection criterion, and identify a cross-modal emotion subspace organized by valence and arousal.
Yi-Bo Wang, Ji-Sheng Dang, Bi-Mei Wang et al.
· 1 citation
· ⚡1