V-Rubrics: 350K Fine-Grained Rubrics Fix Credit Assignment in Multimodal RL

Researchers from NTU S-Lab, A*STAR, and UIUC introduce V-Rubrics , a rubric-based reinforcement learning framework that addresses the credit assignment problem in multimodal models. The paper, accepted at EMNLP 2026, open-sources code, data, and models.

Core Problem: Answer Correctness ≠ Reliable Reasoning

Multimodal models often produce coherent reasoning that nonetheless misreads chart values or hallucinates objects. Outcome-only rewards reinforce these "correct answer, wrong reasoning" trajectories. Conversely, a wrong final answer may still contain valid visual observations that deserve positive signal. V-Rubrics solves this by breaking each response into independently verifiable criteria.

Rubric Design: Three Dimensions, Atomic Criteria

Each rubric checks a single atomic claim and carries an importance tag (Essential, Important, Optional, Pitfall) with weights 1–5 for positive criteria and −1/−2 for Pitfall traps. The three dimensions are:

Visual Faithfulness (VF) — 209,436 criteria (59.3%): every mentioned object, attribute, relation, count, visible text, or chart value must be grounded in the image.

Reasoning Consistency (RC) — 101,369 criteria (28.7%): conclusions must logically follow from the extracted facts.

Instruction Following (IF) — 42,133 criteria (11.9%): format, task requirements, and explicit constraints.

Priority order for visual QA: VF ≫ RC > IF. Rubrics are self-contained; the verifier judges only the generated response against the criteria, without re-reading the original image .

Data Construction: V-Rubrics 50K

Samples drawn from 17 public visual datasets (chart/document QA, diagrams, visual math, counting, educational QA, general visual reasoning). An SFT model (Qwen3-VL-8B-Instruct + OpenMMReasoner-SFT-874K) answered each candidate 8×; questions solved every time were discarded. Remaining 50,248 samples split: hard 36.1%, medium 50.4%, simple 13.6%. Gemini-3-Pro generated rubrics via an image-conditioned protocol, yielding 352,938 total criteria.

Training: Rubric GRPO with Prefix-Located Advantage

Base: Qwen3-VL-8B-Instruct → SFT on OpenMMReasoner-SFT-874K → two RL runs from same checkpoint on identical V-Rubrics 50K data with same sampling budget and optimizer.

Answer-level GRPO : scalar reward = 1 if final answer correct else 0.

Rubric GRPO : semantic reward = 0.5 × answer correctness + 0.5 × rubric satisfaction (per-criterion pass/fail) + format reward.

Key innovation: each rubric’s advantage is computed separately, standardized within the rollout group, and applied only to the response prefix up to the evidence sentence (located by fuzzy matching). If evidence cannot be located, advantage falls back to full response. Pitfall triggers veto: cancels answer credit and positive rubric credit for that response.

Results: Consistent Gains Across Benchmarks

General & Knowledge (Table 1) : Overall Avg. 68.04 vs. 66.25 (answer GRPO) and 64.93 (SFT). MMMU series drives gains: MMMU Val +2.56, MMMU-Pro +2.43, MMMU-Pro Vision +2.60; Knowledge Avg 59.35 → 61.88. MMBench-Dev unchanged (~86.9).

Visual Math, Chart & Logic (Table 2) : Overall Avg. 62.45 vs. 61.94 (answer GRPO) and 58.45 (SFT). MathVision 56.71 → 58.88; LogicVista 60.63 → 62.42; DynaMath Worst +1.20; WeMath Loose +1.43. Math Avg 63.19 → 63.63; Chart Avg 58.81 → 59.51. Qwen3-VL-8B-Thinking reference: 62.22.

Ablation: What Contributes?

On General & Knowledge:

Rubric scoring with scalar sequence-level aggregation: 67.74 (+1.49 over answer GRPO).

Component-wise standardization + prefix-located advantage: 68.04 (+0.30 further).

Thus, finer evaluation content provides most lift; precise credit assignment adds a smaller but measurable gain.

Qualitative Examples

Portrait counting : Answer GRPO correctly detects 4 people but infers one looks >70, concludes born pre-1945, answers 3. Rubric model adds the missing age→birth-year calculation step and answers 4.

Function graph : Answer GRPO misreads first x where f(x)=2 as x=4. Rubric model re-reads the curve segment by segment and returns x=2.

Resources

Paper: https://arxiv.org/abs/2608.25580v1 Project page: https://shulin16.github.io/v-rubrics/ Code: https://github.com/shulin16/v-rubrics Dataset:

https://huggingface.co/datasets/v-rubrics/v-rubrics-50k