一作ACL 2026Poster· Natural Language Processing
LLM - VA : Resolving the Jailbreak-Overrefusal Trade-off via Vector Alignment
Haonan Zhang · Dongxia Wang · Yi Liu · Kexin Chen · Wenhai Wang
摘要
Safety-aligned LLMs suffer from two failure modes: jailbreak (responding to harmful inputs) and over-refusal (declining benign queries). Existing vector steering methods adjust the magnitude of answer vectors, but this creates a fundamental trade-off—reducing jailbreak increases over-refusal and...