Chaymaa Abbas

THESIS-MoE: Trainable Hierarchical Extraction and SteerIng of Sycophancy in Mixture-of-Experts

A shared contrastive signal is introduced that identifies where sycophancy lives across the MoE hierarchy and drives interventions that act only where the behavior is present, demonstrating that sycophancy resides in identifiable computational subcircuits and can be selectively steered while maintaining a favorable rem...