arXivCamila Blank, Agam Bhatia, Senthooran Rajamanoharan, Arthur Conmy, Neel NandaSat, May 30, 2026, 9:22 PM PDT
score 15.9
Hidden traits leak into AI models through hidden steering vectors
Original: Subliminal Learning Is Steering Vector Distillation
Source: arxiv.org ↗
Writing ELI5 summary…