V-Rubrics: rubric-based RL improves visual faithfulness in VLMs
liuziwei7 · x · 2026-08-29
A new paper, 'V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning,' tackles vision-language models that answer fluently but wrongly: a single unsupported object, chart value, or inference step can invalidate an otherwise plausible response.
Problem: The authors frame this as a credit-assignment failure in multimodal post-training — scalar outcome rewards say whether an answer is acceptable, but not which visual facts are grounded or which reasoning steps are valid.
Method:
- Decompose reference responses into atomic propositions, scored on Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF), giving structured partial credit;
- Fine-tune Qwen3-VL-8B-Instruct on OpenMMReasoner-SFT-874K for the SFT starting point;
- Build V-Rubrics 50K: 50,248 examples from 17 visually grounded sources, annotated by Gemini-3-Pro under one protocol;
- Compare two GRPO variants: scalar answer credit vs component-wise, prefix-localized rubric credit.
Results: Rubric-based GRPO beats both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks — showing rubrics are a useful reward abstraction for visual post-training.
More from Models
- Terminal-Bench 4.0: GLM-5.3 Surpasses GPT-5.6 in New Ranking — eyishazyer · 2026-08-29
- Minimax H3 limits: Identity leaking and quality degradation — Kooky-Mode3047 · 2026-08-29
- Qwen3.8 27B runs at 50 tok/s with 100k context on 16GB GPU — qaf23 · 2026-08-29
- Uncensored GLM-5.3-Flash weights released, refusal rates drop to 11% — lipeng0820 · 2026-08-29
- Qwen3.8 QAT Q2 review: Good for quick answers, fails at long context — kirisoraa · 2026-08-29
- Switched to Codex from Claude: Sol fails, loops, and lies — kwesk · 2026-08-29