Diagnosing Visual Reasoning: Challenges, Insights, and a Path Forward

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bi, Jing, Sun, Guangyu, Vosoughi, Ali, Chen, Chen, Xu, Chenliang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912666504658944
author Bi, Jing
Sun, Guangyu
Vosoughi, Ali
Chen, Chen
Xu, Chenliang
author_facet Bi, Jing
Sun, Guangyu
Vosoughi, Ali
Chen, Chen
Xu, Chenliang
contents Multimodal large language models (MLLMs) that integrate visual and textual reasoning leverage chain-of-thought (CoT) prompting to tackle complex visual tasks, yet continue to exhibit visual hallucinations and an over-reliance on textual priors. We present a systematic diagnosis of state-of-the-art vision-language models using a three-stage evaluation framework, uncovering key failure modes. To address these, we propose an agent-based architecture that combines LLM reasoning with lightweight visual modules, enabling fine-grained analysis and iterative refinement of reasoning chains. Our results highlight future visual reasoning models should focus on integrating a broader set of specialized tools for analyzing visual content. Our system achieves significant gains (+10.3 on MMMU, +6.0 on MathVista over a 7B baseline), matching or surpassing much larger models. We will release our framework and evaluation suite to facilitate future research.
format Preprint
id arxiv_https___arxiv_org_abs_2510_20696
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Diagnosing Visual Reasoning: Challenges, Insights, and a Path Forward
Bi, Jing
Sun, Guangyu
Vosoughi, Ali
Chen, Chen
Xu, Chenliang
Computer Vision and Pattern Recognition
Multimodal large language models (MLLMs) that integrate visual and textual reasoning leverage chain-of-thought (CoT) prompting to tackle complex visual tasks, yet continue to exhibit visual hallucinations and an over-reliance on textual priors. We present a systematic diagnosis of state-of-the-art vision-language models using a three-stage evaluation framework, uncovering key failure modes. To address these, we propose an agent-based architecture that combines LLM reasoning with lightweight visual modules, enabling fine-grained analysis and iterative refinement of reasoning chains. Our results highlight future visual reasoning models should focus on integrating a broader set of specialized tools for analyzing visual content. Our system achieves significant gains (+10.3 on MMMU, +6.0 on MathVista over a 7B baseline), matching or surpassing much larger models. We will release our framework and evaluation suite to facilitate future research.
title Diagnosing Visual Reasoning: Challenges, Insights, and a Path Forward
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2510.20696