Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination
The paper presents findings on modality-conflict hallucination in multimodal large language models (MLLMs). It identifies an imbalance in attention heads that either drive or resist hallucination during generation. The authors propose a new intervention method, MACI, which effectively reduces hallucinations while maintaining accuracy.
- ▪Modality-conflict hallucination occurs when MLLMs prioritize incorrect textual information over visual evidence.
- ▪The study reveals two groups of attention heads: those that drive hallucinations and those that resist them.
- ▪The proposed MACI intervention suppresses driving heads when conflicts are detected, leading to significant reductions in hallucinations.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.19250 (cs) [Submitted on 19 May 2026] Title:Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination Authors:Jinrui Jiang, Zhangtai Wu, Zhen Wu, Xinyu Dai View a PDF of the paper titled Causal Evidence for Attention Head Imbalance in Modality Conflict Hallucination, by Jinrui Jiang and 2 other authors View PDF HTML (experimental) Abstract:Modality-conflict hallucination occurs when multimodal large language models (MLLMs) prioritize erroneous textual premises over contradictory visual evidence. To understand why visual evidence fails to prevail during generation, we take a mechanistic perspective and examine which internal components drive or resist this failure.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
