arxivcs.CV2026-07-31
Role-Break in Attention Heads: Understanding and Detecting Hallucinations in VLMs
Mingyu Wang, Weilin Jin, Wenbo Li, Haoyang Huang, Nan Duan, Tong Jia, et al.
Despite remarkable progress in vision-language generation, Vision-Language Models (VLMs) remain prone to hallucinations, producing content that is inconsistent with or unsupported by the input image. Existing works largely design detection or mitigation methods around one specifi…