Multimodal Large Language Models (MLLMs) frequently hallucinate due to their reliance on fragile, linear reasoning and weak visual grounding. We propose Visual Attention Reasoning (VAR), a reinforcement learning framework that reformulates reasoning as a hierarchical search with self-verification. VAR enforces traceable evidence grounding by generating explicit bounding boxes, guided by a novel reward function combining geometric precision and semantic sufficiency. Furthermore, it replaces linear Chain-of-Thought with a tree-search policy capable of backtracking to correct logical errors. Theoretical analysis validates the framework's reliability, and extensive experiments demonstrate that VAR significantly outperforms state-of-the-art methods on complex hallucination and safety benchmarks.
@article{arxiv.2510.18619,
title = {Visual Attention Reasoning via Hierarchical Search and Self-Verification},
author = {Wei Cai and Jian Zhao and Yuchen Yuan and Tianle Zhang and Ming Zhu and Haichuan Tang and Xuelong Li},
journal= {arXiv preprint arXiv:2510.18619},
year = {2026}
}
Comments
The paper is withdrawn by the authors after discovering a flaw in the theoretical derivation presented in the Method section. This incorrect step leads to conclusions that are not supported by the corrected derivation. The authors plan to reconstruct the argument and will release an updated version once the issue is fully resolved