Open-o3-Video-v2 grounds video reasoning with verifiable spatio-temporal evidence
PKU · hf · 2026-10-14
Open-o3-Video-v2 (PKU, Fudan, NTU, UC Merced) frames video reasoning as a unified agentic evidence-grounded process: the model actively identifies missing information, acquires extra observations, and grounds timestamps and object-level bounding boxes into its reasoning trace. The team builds STGR-v2-CoT135K (52K multi-turn grounded trajectories) plus STGR-v2-RL-13K for RL, and proposes Clue-Seek GRPO (CS-GRPO), rewarding clue coverage and penalizing redundant retrieval. The 4B model hits 40.4% mAM and 56.3% mLGM on V-STaR, beating its Qwen3-VL-4B backbone by 8.5% and 12.6%. Code and models are open-sourced.
More from Multimodal
- Prompt trick: expand a headshot into a half-body portrait, head intact — sujingshen · 2026-10-14
- One image plus one prompt makes a complete horror micro-film — umesh_ai · 2026-10-14
- Rick & Morty-style promo video marks 36th style in huashu-art-motion AI project — AlchainHust · 2026-10-14
- Drama 3 35B multi-speaker TTS gets int4 quantization: 22.1GB, runs on one GPU — Illustrious_Row_9971 · 2026-10-14
- Claude Code makes a SpaceX documentary video entirely via chat, no editing software — xiaohu · 2026-10-14
- Four made-up tokens unlock surreal scenes in Midjourney V8.2 — LudovicCreator · 2026-10-14