Bridging the Gap Between Harmfulness Belief and Refusal Behavior for Safety Alignment
Published in NeurIPS, 2026
Recommended citation: Zhang, L., Feng, C., Wang, Q., Shen, W., & Wei, Z. Bridging the Gap Between Harmfulness Belief and Refusal Behavior for Safety Alignment. In NeurIPS 2026.
Abstract. This paper proposes Bridging Harmfulness and Refusal (BHR), a training framework that carries harmfulness information from the instruction to the response-start representation to improve jailbreak robustness while reducing over-refusal and preserving general capabilities.
Authors: Lu Zhang*, Chen Feng*, Qingzhuo Wang, Wen Shen†, Zhihua Wei†.
* Equal contribution; † Corresponding author.