Title: Mitigating Hallucinations by Rectifying Visual Blur Caused by Attention Diversion: A Theoretical and Algorithmic Study

Abstract: While multimodal large language models (MLLMs) are prone to object hallucinations, the visual perceptual mechanisms driving this issue remain largely unclear. This study demonstrates that such hallucinations are closely linked to a phenomenon analogous to human attention distraction. In humans, divided focus leads to diminished visual acuity and erroneous descriptions; similarly, in AI models, this manifests as spatial irregularities in multi-head attention and a temporal decay of attention assigned to image tokens during the decoding phase. Our theoretical analysis indicates that this dispersion of attention heightens model complexity and impairs the generalization capability of classification tasks. Based on these insights, we introduce the Attention-Focused Approach for Improved Image Perception (AFIP). This method addresses attention diversion by enriching cross-head attention and strengthens visual grounding via dynamic enhancement of historical attention. Comprehensive experiments across various benchmarks and models confirm that AFIP effectively reduces hallucinations without requiring additional training.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC