AI RESEARCH

DSPA: Dynamic SAE Steering for Data-Efficient Preference Alignment

arXiv CS.AI

ArXi:2603.21461v1 Announce Type: cross Preference alignment is usually achieved by weight-updating