Visual Rubrics · Multimodal Post-Training

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

Shulin Tian1,2,* Minglun Li3,* Yuhao Dong1,‡,* Hao Ding3,* Jiarui Yao4 Haiwen Diao1 Jingkang Yang1 Hongyuan Zhu2 Ziwei Liu1
1S-Lab, Nanyang Technological University 2A*STAR 3Independent Researcher 4UIUC

*Equal Contributions ‡Project Lead

shulin002@e.ntu.edu.sg

Code Hugging Face Dataset BibTeX

Fine-grained multimodal feedback

Fluent answers can still be visually wrong.

V-Rubrics turns each response into atomic checks for visual grounding, reasoning, and instruction following.

VFVisual Faithfulness
RCReasoning Consistency
IFInstruction Following
A spring-tide visual question decomposed into six weighted visual-faithfulness, reasoning-consistency, and instruction-following rubric items.
One response, three dimensions, independently checkable credit.
50,248
training examples
17
visual sources
3
rubric dimensions
10
benchmark families

Abstract

Vision-language models can produce fluent answers that are visually wrong: a single unsupported object, chart value, or intermediate inference can invalidate an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards say whether an answer is acceptable, but not which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain a common SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner’s cold-start data recipe. This checkpoint drives rejection sampling during the construction of V-Rubrics 50K, a 50K-example training set from 17 visually grounded sources, and initializes two GRPO variants using either scalar answer credit or component-wise, prefix-localized rubric credit. Rubric-based GRPO improves over both the shared SFT baseline and answer-level GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.

Dataset

V-Rubrics 50K

50,248 rubric-annotated examples retained from 85,217 candidates across 17 public visual datasets.

V-Rubrics 50K distribution: 59.1 percent visual-faithfulness, 28.9 percent reasoning-consistency, and 11.9 percent instruction-following rubrics, with the retained source distribution.
Rubric dimensions inside; retained source proportions outside.
Selected candidates85,217drawn from 17 sources
Passed rule checks57,17267.1% of candidates
Final after SFT sampling50,24859.0% of candidates
Stage 1

Remove invalid image–instruction–reference records.

Stage 2

Balance hard, medium, and simple curriculum bins from the rejection-sampling signal.

18,121
hard · 36.1% · 0/8
25,306
medium · 50.4% · 1/8–5/8
6,821
simple · 13.6% · 6/8–8/8
Source-by-source counts: 85,217 selected candidates → 50,248 final
Per-source examples selected and retained for V-Rubrics 50K
Source datasetSelected candidates*Passed checksFinalFinal / selected
AI2D2,0542,0541,75885.6%
ChartGalaxy2,9001,5001,28944.4%
ChartQA76076068790.4%
ChartQA-X4,1451,5001,31431.7%
ChartX1,2741,2741,20394.4%
CharXiv74674668191.3%
DocVQA37137133690.6%
Geometry3K1,0741,07491985.6%
InfographicVQA4,0342,5002,07151.3%
MathVerse2,9422,2001,29043.8%
MM-K126,4046,3006,15396.1%
PixMo-Count5,1322,5002,12141.3%
ThinkLite-Hard9,2269,2268,62493.5%
ThinkLite33,40915,00013,04139.0%
ViRL39K7,3047,3046,18884.7%
We-Math2.0-STD1,2631,2631,07685.2%
We-Math2.0-Pro2,1791,6001,49768.7%
Total85,21757,17250,24859.0%

* Candidate counts are pipeline selections, not full upstream dataset sizes.

Release contents and terms. The companion dataset packages the selected records, including image payloads, rubric annotations, and provenance metadata. Image and upstream-record rights remain source-dependent; inclusion in V-Rubrics does not relicense them, and the repository's Apache-2.0 license applies only to first-party software.

Our method

From visual data to rubric-guided RL

Full two-panel V-Rubrics pipeline: candidate visual data is filtered and sampled before atomic VF, RC, and IF rubrics form V-Rubrics 50K. A shared SFT checkpoint then initializes the actor and frozen KL reference; grouped responses receive answer credit across the full response and item-wise rubric credit through evidence-prefix masks before GRPO updates.
Left: candidates from 17 visual datasets are filtered, difficulty-balanced, and converted into atomic VF/RC/IF rubrics to form V-Rubrics 50K. Right: a shared SFT policy generates grouped rollouts, while answer-level and prefix-localized rubric credit guide GRPO. The two signals are mixed equally (α = 0.5).

Performance

Performance Results

The full comparison reports all 22 models from the paper, with detailed panels for the five checkpoints evaluated under our setup.

22 models 16 metrics + aggregates 4 model groups
+1.79over answer GRPO
general + knowledge
+4.00over SFT
visual reasoning
+0.30component + prefix over scalar rubric credit
67.74 → 68.04
Overview

Matched aggregate scores across all benchmark panels.

GENERAL + KNOWLEDGE

Overall average

Qwen3 Instruct
67.63
Qwen3 Thinking
69.55
SFT
64.93
Answer GRPO
66.25
Rubric GRPO
68.04
VISUAL MATH + CHART + LOGIC

Overall average

Qwen3 Instruct
58.56
Qwen3 Thinking
62.22
SFT
58.45
Answer GRPO
61.94
Rubric GRPO
62.45

Shared panel axis: 55–70. Exact values are printed beside every bar.

General + knowledge

MMBench-Dev and the knowledge-oriented MMMU family.

MMBench-Dev

Qwen3 Instruct
86.08
Qwen3 Thinking
87.29
SFT
84.79
Answer GRPO
86.94
Rubric GRPO
86.51

MMMU Val

Qwen3 Instruct
69.00
Qwen3 Thinking
72.22
SFT
66.78
Answer GRPO
68.00
Rubric GRPO
70.56

MMMU-Pro · 10c/Std

Qwen3 Instruct
57.75
Qwen3 Thinking
59.48
SFT
54.34
Answer GRPO
55.72
Rubric GRPO
58.15

MMMU-Pro V

Qwen3 Instruct
57.69
Qwen3 Thinking
59.19
SFT
53.82
Answer GRPO
54.34
Rubric GRPO
56.94

Knowledge Avg.Qwen3 Instruct 61.48Qwen3 Thinking 63.63SFT 58.31Answer GRPO 59.35Rubric GRPO 61.88

Shared panel axis: 50–90. Exact values are printed beside every bar.

Visual math

Five visual-mathematical and grounded reasoning benchmarks.

MathVista mini

Qwen3 Instruct
76.60
Qwen3 Thinking
77.80
SFT
78.30
Answer GRPO
81.10
Rubric GRPO
81.30

MathVision test

Qwen3 Instruct
56.41
Qwen3 Thinking
62.70
SFT
55.46
Answer GRPO
56.71
Rubric GRPO
58.88

MathVerse V/O

Qwen3 Instruct
47.97
Qwen3 Thinking
52.03
SFT
47.84
Answer GRPO
52.16
Rubric GRPO
49.37

DynaMath Worst

Qwen3 Instruct
40.72
Qwen3 Thinking
40.32
SFT
41.32
Answer GRPO
41.12
Rubric GRPO
42.32

WeMath Loose

Qwen3 Instruct
75.81
Qwen3 Thinking
84.67
SFT
77.43
Answer GRPO
84.86
Rubric GRPO
86.29

Math Avg.Qwen3 Instruct 59.50Qwen3 Thinking 63.50SFT 60.07Answer GRPO 63.19Rubric GRPO 63.63

Shared panel axis: 40–90. Exact values are printed beside every bar.

Chart + logic

LogicVista and CharXiv reasoning.

LogicVista

Qwen3 Instruct
62.19
Qwen3 Thinking
63.53
SFT
60.63
Answer GRPO
60.63
Rubric GRPO
62.42

CharXiv reasoning

Qwen3 Instruct
50.20
Qwen3 Thinking
54.50
SFT
48.20
Answer GRPO
57.00
Rubric GRPO
56.60

Chart Avg.Qwen3 Instruct 56.20Qwen3 Thinking 59.02SFT 54.42Answer GRPO 58.81Rubric GRPO 59.51

Shared panel axis: 45–65. Exact values are printed beside every bar.

Full model comparison

Select any paper metric to rank all 22 models; results not reported in the paper are shown as “-”.

Closed-source Open-source Instruct Open-source Thinking Ours Evaluated badges mark checkpoints run under our evaluation setup; other comparison values follow the cited paper sources.

Largest rubric gains: MMMU, MMMU-Pro, MathVision, DynaMath, WeMath Loose, and LogicVista. Answer GRPO remains higher: MMBench-Dev, MathVerse V/O, and CharXiv reasoning.

Qualitative analysis

Local feedback changes the reasoning path

Two qualitative comparisons between answer-only GRPO and rubric-trained checkpoints: a people-counting age-inference example and a piecewise graph-tracing example.
Rubric training corrects the unsupported age inference (top) and follows the visible graph segments (bottom). The first example remains visually ambiguous; these cases are illustrative only.

Limitations

Structured feedback is still model-generated feedback

01

Generated rubrics may inherit answer bias, ambiguity, or unsupported assumptions.

02

Fuzzy span matching is not exact token-level supervision.

03

Benchmark accuracy only approximates visual faithfulness.

04

Qwen-family judges may favor Qwen-family policies.

Citation

BibTeX

If you use V-Rubrics, please cite:

V-Rubrics · 2026 
@misc{tian2026vrubrics,
  title  = {V-Rubrics: Visual Faithfulness via Rubric-Based
            Reinforcement Learning},
  author = {Shulin Tian and Minglun Li and Yuhao Dong and
            Hao Ding and Jiarui Yao and Haiwen Diao and
            Jingkang Yang and Hongyuan Zhu and Ziwei Liu},
  year   = {2026},
  url    = {https://shulin16.github.io/v-rubrics/}
}