ByteDance's Seed team open-sources the entire UniVR-34B series weights + VR-X dataset + training code under CC BY 4.0 on Hugging Face, proposing the first framework that simultaneously learns complex reasoning, fine-grained physical dynamics, and long-horizon planning directly from pure visual demonstrations. Unlike the mainstream CoT path that goes through textual intermediate steps, UniVR directly does next-token prediction in the visual space, using image sequences as the carrier of thought. The architecture is based on BAAI Emu3.5 (a 34B VQ-VAE unified generation model), and training uses SFT (310k samples) → VR-GRPO RL (3k samples) in two stages. The most worth-mentioning is VR-GRPO's reward design: traditional GRPO only uses a global reward to evaluate task completion, which is easily reward-hacked in long-horizon tasks; VR-GRPO simultaneously gives a Step-Focal reward — using the variance of CLIP features between rollout samples to locate the sub-step most prone to failure, then a VLM evaluator for fine-grained scoring. The final reward is designed as R_reason = R_g − λ|R_g − R_s|, constraining both terminal correctness and intermediate physical compliance and logical coherence. On the self-built VR-X Benchmark (16 sources, 6 task types, 1.8k evaluations), UniVR 34B gets 58.2 Overall, an 18.4pp lift over the Emu3.5 baseline (39.8), with Robot tasks +25.2, Editing +15.8, Spatial +11.2. Horizontally, it's only 8pp behind Gemini 3 Pro + Nano Banana 2 (66.1), but UniVR is only 34B — half the parameter count, catching up to the closed-source multimodal combination. At the same time it gains on six general multimodal benchmarks (MMMU / MME / MMBench / MathVista / MM-Vet), with zero main-task capability regression. I think the industry signal here is: visual space is becoming the new "thinking medium". For robot manipulation, tool calling, long-horizon planning, pressing the reasoning trace into the pixel layer is more natural than the token layer — physical consistency, spatial relationships, action bridging — these dimensions take a long detour to express in text CoT, while images are inherently continuous. At the same time, VR-GRPO's dual-reward structure also provides a general template for long-horizon Agent RL: don't just watch task success or failure, watch the steps most likely to break. ByteDance's move pulls Emu3.5's generative foundation, the Seed series of video generation, and UniVR's visual planning onto the same representation axis — whether the next step can integrate video generation of the Sora 2 / Gen-4.5 tier into the same visual-reasoning framework is worth watching.