ByteDance's UniVR Thinks in Pictures, Not Words. Its Chain of Thought Is Generated, Frame by Frame.

TL;DR

ByteDance and Beijing Jiaotong University quietly dropped UniVR, a 34B open-weight model that throws out the usual playbook: it does not reason in text. Instead of writing a chain of thought in words, it generates one as a sequence of images, sketching the intermediate steps of a task the way you would rough out a plan on a whiteboard. Built on Emu3.5 and trained on the team's new VR-X data, it lifts that base model's score on the VR-X reasoning benchmark from 39.8 to 58.2, and it ships under a permissive CC BY 4.0 license.

Text models reason in words. UniVR reasons in pixels.

For two years, "reasoning" has meant one thing: the model writes out its thinking in text before it answers. That works beautifully for math and code, and poorly for anything words cannot easily hold. Ask a text reasoner to plan a robot's grasp, edit a photo step by step, or navigate a maze, and it has to translate a visual problem into sentences, reason over the sentences, and translate back. Detail leaks at every hop.

UniVR skips the round trip. Given an image and an instruction, it generates the next frames of the solution directly, in pixel space, using a single next-token objective over visual tokens. The reasoning trace is not a paragraph describing what should happen. It is the pictures of it happening.

A text-only model that "shows its work" on a visual task is really narrating a picture it cannot see. UniVR just draws the picture, the way a chess player thinks by imagining the board three moves ahead rather than reciting the moves aloud.

How it works

Under the hood, UniVR keeps the 34B backbone of BAAI's Emu3.5, a unified generative model that already treats vision and language as one long stream of tokens. What changes is what the model spends its generation budget on: coherent reasoning frames, not just a final answer.

The training recipe

The dataset, VR-X, is the quiet workhorse: 1.5M raw samples pulled from 16 sources, distilled to 310k supervised examples and 3k reinforcement-learning prompts, spanning six task families (visual guidance, robotic manipulation, editing, spatial perception, visual search, and puzzle games).

Training runs in two stages. A supervised cold start first teaches the model to produce coherent visual traces at all. Then a reinforcement-learning phase the authors call VR-GRPO tunes it with two reward signals at once: a global reward that grades the whole trajectory, and a step-focal reward that zeroes in on the error-prone moments where a plan usually falls apart.

The numbers

On VR-X the jump is large: the overall score climbs from Emu3.5's 39.8 to 58.2. Because the team built VR-X, treat that as an in-house result, not an independent verdict. The more reassuring signal is that general multimodal ability did not regress to buy it. On MMMU the score edges from 0.292 to 0.337, and on MathVista from 41.7 to 44.0. Small gains, but they mean the visual-reasoning training added a skill without knocking out the old ones, which is the failure mode everyone worries about.

Why builders should care

Two reasons. First, the weights are actually open. At 34B under CC BY 4.0, this is something you can download, run, and fine-tune, not a demo behind a waitlist. For anyone working on robotics, UI agents, image-editing pipelines, or spatial planning, a model that reasons in the same modality as the task is worth a serious look.

Second, it is a real data point in an argument that has mostly been hand-waving: that text is a lossy container for some kinds of thought. UniVR does not prove text reasoning is dead. It shows that for visual tasks, letting the model think in pictures beats forcing it to think in captions, and it hands you the weights to check the claim yourself.

The caveats

Keep expectations calibrated. The headline gains sit on the authors' own benchmark, scored in part by a VLM judge, so the 39.8 to 58.2 figure needs outside replication before you bank on it. The gains on independent benchmarks are modest. This is a 34B research release, not a product: you need real GPUs to run it, the paper's arXiv link is still a placeholder on the project page, and the linked code repository was returning a 404 at publication time, so for now the weights on Hugging Face are the usable artifact. Treat UniVR as an early, promising instance of a bigger idea, not a finished tool.

Key Takeaways

• UniVR is a 34B open-weight (CC BY 4.0) model from ByteDance and Beijing Jiaotong University that reasons by generating images, not text.
• Given an image and an instruction, it produces a "visual reasoning trace," the next frames of the solution, via a single next-token objective over visual tokens.
• It is built on BAAI's Emu3.5 and trained on the new VR-X data (1.5M raw samples, six task families) with a supervised cold start plus VR-GRPO reinforcement learning.
• On the team's own VR-X benchmark it lifts the base model from 39.8 to 58.2, an 18-point gain, while MMMU and MathVista nudge up rather than regress.
• The bet: for visual tasks like robotics, editing, and spatial planning, thinking in pixels beats translating the problem into text and back.
• Caveats: the big gains are on an in-house, VLM-judged benchmark, independent-benchmark gains are modest, and this is a research-grade 34B release, not a polished product.

Sources: UniVR-34B-Planning model card (Hugging Face), UniVR project page, Emu3.5 (arXiv 2510.26583).