UniVR: A New Approach to Pure Visual Reasoning
ByteDance · hf · 2026-07-17
ByteDance proposed UniVR, exploring how to learn broader world knowledge relying solely on pure visual demonstrations, while simultaneously covering three capabilities:
- Complex reasoning
- Fine-grained physical dynamics modeling
- Long-term planning
The core method is VR-GRPO: a reinforcement learning paradigm combining global and step-level rewards to constrain logical and physical consistency during reasoning, without relying on task-specific heuristics or image-text paired data.
They also built the VR-X benchmark, aggregating 16 sources to cover long-horizon manipulation, spatial puzzles, and physical reasoning, serving as a comprehensive evaluation suite under a pure visual protocol. Experiments show UniVR improves performance on VR-X by up to 25%, and its enhanced visual reasoning capabilities also feed back into multiple multimodal understanding benchmarks. Code, data, and models are all open-sourced.
More from Multimodal
- LichtFeld Studio v0.5.4.3 adds community-built SteadySplats and voxelization features — janusch_patas · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26
- Claude Can Replace Your YouTube Thumbnail Designer in 10 Minutes — thedankoe · 2026-10-26
- Multimodal AI Is Not Useless: A Practitioner Pushes Back on Growing Skepticism — thoma_gu · 2026-10-26
- One prompt, full Blender animation: dev says he never touched Blender once — iamfakhrealam · 2026-10-26
- Dense instance segmentation via Muse Wonder + SAM 3.1 on Meta's Model API — nikhilaravi · 2026-10-26