: Multimodal Judgment via Grounded Verification

Bhavesh Kumar, Dylan Feng, Leonard Tang · 2026

Read : Multimodal Judgment via Grounded Verification by Bhavesh Kumar, Dylan Feng, Leonard Tang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-07990.txt.

Description

Kumar and colleagues train MJ1, a multimodal judge that grounds its decisions in visual evidence. It uses a grounded verification chain running from observations through claims and verification to evaluation and scoring. A counterfactual consistency reward penalizes position bias, keeping the judge honest about where evidence sits. Even without training, the mechanism lifts base-model accuracy on MMRB2. After training, MJ1, with only 3 billion active parameters, reaches 77 percent accuracy on MMRB2 and beats far larger models like Gemini-3-Pro.

Catalog record

CREATORBhavesh Kumar, Dylan Feng, Leonard Tang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31461
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.07990

Browse the full library · About the Tower of Babel