UniVR: Thinking in Visual Space for Unified Visual Reasoning

Researchers at Beijing Jiaotong University and ByteDance introduce UniVR, a unified autoregressive model that learns reasoning and planning directly in visual space rather than converting visual states into language. Built from Emu3.5-34B, UniVR uses VR-GRPO, a reinforcement-learning method combining a global task-completion reward with a Step-Focal reward that samples high-uncertainty substeps using CLIP feature variance, targeting logical gaps and physical violations in long trajectories. The team creates VR-X from 1.5M raw samples across 16 sources, retaining 310k cold-start examples, 3k reinforcement-learning examples, and 1.8k evaluation trajectories covering manipulation, cooking, navigation, puzzles, search, editing, and spatial reasoning. On VR-X, UniVR reaches an overall score of 58.2, versus 39.8 for Emu3.5, an 18.4% gain, with improvements reaching 25% across individual categories. It also improves multimodal understanding, raising MMMU from 0.292 to 0.337 and MM-Vet from 28.0 to 35.6. At 34B parameters, UniVR surpasses or approaches pipelines such as Google DeepMind’s Gemini 3 Pro with Nano Banana 2 and competes favorably with OpenAI’s GPT-5, demonstrating that raw visual demonstrations can teach long-horizon policies without dense image-text supervision.