Beyond Token-Level Guidance: Inference-Time Alignment of Specialized LLMs via Cross-Family Representation Steering
CREST is proposed, an inference-time alignment method that steers base model hidden representations using safety directions extracted from a guidance model of any family, avoiding token-level structural limitations entirely and outperforming baselines by up to 22.2\% on safety benchmarks.
Jin Gan, Xin Li, Jun Luo
· 0 citations