V-Rubrics: 50k visual samples split into 353k checkable criteria fix multimodal RL credit assignment
jiqizhixin · x · 2026-10-04
NTU S-Lab, ASTAR, and UIUC present V-Rubrics, tackling credit assignment in multimodal RL: models can misread a chart number, hallucinate an object, and still land the correct final answer—result-scored RL rewards it anyway; conversely, a wrong answer doesn't mean every observation was wrong.
Key approach:
- Decomposes 50,248 visual samples into 352,938 individually checkable criteria;
- Visual facts, reasoning steps, and task requirements each participate separately in reward computation;
- Each criterion's score feeds GRPO reinforcement learning, reinforcing what was right and correcting what was not.
More from Multimodal
- TanStack AI adds experimental text-to-GLB 3D model and prompt-to-motion animation generation — bennash · 2026-11-20
- The Tide Machine: a generative art piece concepted by a human and orchestrated by GPT Astra 6.2 — dreamwieber · 2026-11-20
- Minimax H3 Ref2Va + VHS LoRA: recreating 90s videotape aesthetics — Engine_Hungry · 2026-11-20
- Ideogram 4.6 released, and it's the open-weights version — skipping 4.5 — KokaOP · 2026-11-20
- Google opens Project Genie 4 waitlist access to Google AI Pro subscribers — testingcatalog · 2026-11-20
- Tesseract brings 3D spatial audio generation to Claude and ChatGPT agents — Ror_Fly · 2026-11-20