V-Rubrics: Rubric-Based RL Makes Multimodal Reasoning Visually Faithful

青稞AI · wechat · 2026-10-06

A team from NTU S-Lab, ASTAR and UIUC proposes V-Rubrics, tackling the problem that vision-language models can produce fluent answers while misreading chart values or citing conclusions unsupported by the image.

The idea refines credit assignment: visual facts, reasoning steps and task requirements become separate rubric criteria, whose per-item scores feed GRPO reinforcement learning so reasoning stays faithful to visual evidence. Accepted to EMNLP 2026 main conference; code and the V-Rubrics-50K dataset are open-sourced. Author Tian Shulin (NTU PhD student) will give a live talk on Oct 8 covering dataset construction, per-item rewards and prefix credit assignment design.

Related event: V-Rubrics tackles multimodal RL credit assignment with 350K fine-grained criteria(3 posts)→

Original post →

More from Multimodal

Multimodal channel →