2026-08-26 Daily Report: V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning

V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning

Authors Shulin Tian, Minglun Li, Yuhao Dong, Hao Ding, Jiarui Yao, Haiwen Diao, Jingkang Yang, Hongyuan Zhu, Ziwei Liu

Affiliations Nanyang Technological University / University of Illinois Urbana-Champaign / A*STAR / Independent Researcher

Categories Method / Reinforcement Learning / Visuals-based reinforcement learning approach, Evaluation / Visual Faithfulness / Assessing adequacy of visual evidence, Application / Model Training / Training with rubric credit signals

License CC BY 4.0

Abstract Overview

This paper studies visual faithfulness in vision-language model post-training and frames the problem as a credit-assignment issue rather than only an evaluation issue. The authors introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic rubric items spanning visual faithfulness, reasoning consistency, and instruction following. They also build V-Rubrics 50K, a 50,248-example training set drawn from 17 visually grounded sources and annotated with weighted rubric criteria. Starting from a Qwen3-VL-8B-Instruct model fine-tuned on OpenMMReasoner-SFT-874K, they train with GRPO using component-wise and prefix-localized rubric credit. Experiments compare this approach against the shared SFT baseline and answer-only GRPO across general, knowledge, visual-math, chart, and logic benchmarks.

Novelty

The paper's main novelty is using visually grounded, instance-specific rubrics as reinforcement learning reward components for VLM post-training, instead of relying only on sequence-level answer rewards. It is also distinctive in combining item-level rubric scoring with prefix-localized credit assignment and in releasing the V-Rubrics 50K dataset for this training setup.

Results

Rubric-based GRPO outperforms both the shared SFT baseline and answer-only GRPO overall, with the clearest gains on reasoning-heavy and visually grounded tasks. The paper reports an Overall Avg. of 68.04 on the general/knowledge suite versus 64.93 for SFT and 66.25 for answer-only GRPO, and 62.45 on the visual-math/chart/logic suite versus 58.45 and 61.94 respectively. Ablation results also favor the full component-wise, prefix-localized rubric design over scalar sequence-level rubric aggregation.

Key Points

V-Rubrics converts reference answers into atomic weighted criteria across visual faithfulness, reasoning consistency, and instruction following, enabling finer-grained reward signals.
The authors construct V-Rubrics 50K from 17 public visually grounded data sources, with 50,248 examples and 352,938 rubric items used for RL training.
Empirical gains are concentrated on benchmarks that depend on grounded intermediate reasoning, while general VLM capability is largely preserved rather than uniformly improved across all tasks.

References

arXiv: https://arxiv.org/abs/2608.25580v1
Fugu-MT: https://fugumt.com/fugumt/paper_check/2608.25580v1
Hugging Face Papers: https://huggingface.co/papers/2608.25580
Project: https://shulin16.github.io/v-rubrics/

Project