Research

Paper

TESTING March 09, 2026

MJ1: Multimodal Judgment via Grounded Verification

Authors

Bhavesh Kumar, Dylan Feng, Leonard Tang

Abstract

Multimodal judges struggle to ground decisions in visual evidence. We present MJ1, a multimodal judge trained with reinforcement learning that enforces visual grounding through a structured grounded verification chain (observations $\rightarrow$ claims $\rightarrow$ verification $\rightarrow$ evaluation $\rightarrow$ scoring) and a counterfactual consistency reward that penalizes position bias. Even without training, our mechanism improves base-model accuracy on MMRB2 by +3.8 points on Image Editing and +1.7 on Multimodal Reasoning. After training, MJ1, with only 3B active parameters, achieves 77.0% accuracy on MMRB2 and surpasses orders-of-magnitude larger models like Gemini-3-Pro. These results show that grounded verification and consistency-based training substantially improve multimodal judgment without increasing model scale.

Metadata

arXiv ID: 2603.07990
Provider: ARXIV
Primary Category: cs.LG
Published: 2026-03-09
Fetched: 2026-03-10 05:43

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.07990v1</id>\n    <title>MJ1: Multimodal Judgment via Grounded Verification</title>\n    <updated>2026-03-09T05:55:48Z</updated>\n    <link href='https://arxiv.org/abs/2603.07990v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.07990v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Multimodal judges struggle to ground decisions in visual evidence. We present MJ1, a multimodal judge trained with reinforcement learning that enforces visual grounding through a structured grounded verification chain (observations $\\rightarrow$ claims $\\rightarrow$ verification $\\rightarrow$ evaluation $\\rightarrow$ scoring) and a counterfactual consistency reward that penalizes position bias. Even without training, our mechanism improves base-model accuracy on MMRB2 by +3.8 points on Image Editing and +1.7 on Multimodal Reasoning. After training, MJ1, with only 3B active parameters, achieves 77.0% accuracy on MMRB2 and surpasses orders-of-magnitude larger models like Gemini-3-Pro. These results show that grounded verification and consistency-based training substantially improve multimodal judgment without increasing model scale.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <published>2026-03-09T05:55:48Z</published>\n    <arxiv:primary_category term='cs.LG'/>\n    <author>\n      <name>Bhavesh Kumar</name>\n    </author>\n    <author>\n      <name>Dylan Feng</name>\n    </author>\n    <author>\n      <name>Leonard Tang</name>\n    </author>\n  </entry>"
}