V-Rubrics: NTU team uses 352K rubric criteria to fix credit assignment in visual RL
青稞AI · wechat · 2026-10-05
Researchers from NTU S-Lab, ASTAR and UIUC propose V-Rubrics, a rubric-based reinforcement learning approach for credit assignment in multimodal reasoning, accepted to EMNLP 2026 main conference with code and data open-sourced.
- Answers are scored along three dimensions: Visual Faithfulness (VF), Reasoning Consistency (RC) and Instruction Following (IF), yielding 352,938 atomic criteria over 50,248 samples in V-Rubrics 50K (VF accounts for 59.3%).
- The verifier reads only the response and self-contained criteria (not the raw image); each rubric's advantage is localized to the response prefix up to its evidence sentence, and Pitfall items can veto rewards for hallucinations.
- From the same Qwen3-VL-8B SFT checkpoint and identical RL data, rubricGRPO reaches 68.04 OverallAvg. on general/knowledge benchmarks vs 66.25 for answer-level GRPO, and 62.45 on visual math/chart benchmarks, with gains concentrated on MMMU series.
More from Multimodal
- Anima Base Workflow Produces Only Static Noise With 3 Models: Debugging Help — Casualoldgamer1 · 2026-11-20
- Fine-tuned Qwen3-VL 8B reads camera shots at 69% accuracy, beating Opus and Gemini — ravisparikh · 2026-11-20
- Redditor Shares 'Entrega Express', a Short Film Made With Seedance 2 — Ok-Nerve941 · 2026-11-20
- HeyGen Releases Avatar 6, Shows Side-by-Side Realism Leap Since Avatar III — HeyGen · 2026-11-20
- User calls Google Genie's new upgrade impressive, diving back in ASAP — AI_Andrew · 2026-11-20
- 'Johnny' — stunning Johnny Depp-style AI portraits showcase top image generation quality — egeberkina · 2026-11-20