一作ACL 2026Findings· Findings of ACL
ECHA : Jailbreaking LVLM s via the Mismatch between Implicit Semantic Reconstruction and Explicit Safety Alignment
Chenxing Xu · Junyong Jiang · Zehu Zhang · Lu Dong
摘要
Large Visual Language Models (LVLMs) achieve superior multimodal reasoning but inevitably expand the safety attack surface. While recent studies have explored emoji-based vulnerabilities, they predominantly focus on textual tokenization artifacts and neglect the model’s intrinsic capability to in...