Conceptrefusal alignment
Refusal Alignment
A behavioral alignment mechanism in instruction-tuned language models that resists harmful requests. It relies on a specific refusal direction in representation space and is an important component of model safety.
Timeline (last 90 days)
Nov 5
Encoder-based context compression systematically weakens refusal behavior in instruction-tuned language models, despite high reconstruction fidelity.
Unverified50%