Attention-space Contrastive Guidance for Efficient Hallucination Mitigation in LVLMs

ArXi:2601.13707v2 Announce Type: replace-cross Hallucinations in large vision--language models (LVLMs) often arise when language priors dominate over visual evidence, leading to object misidentification and visually inconsistent descriptions. We address this problem by framing hallucination mitigation as contrastive guidance that steers generation toward visually grounded and semantically faithful text. We propose Attention-space Contrastive Guidance (ACG), a