UniVR: Thinking in Visual Space for Unified Visual Reasoning

Zhongwei Ren, Yunchao Wei, Yao Zhao, Weibo Gong, Xiao Liu, Anran Wang, Xiangtai Li, Xiaojie Jin

UniVR enables complex reasoning and planning directly in visual space using a reinforcement learning paradigm with step-level rewards.

How can we enable models to perform complex reasoning and physical planning natively in visual space rather than relying on text-based chain-of-thought?

Current AI models rely on textual abstractions to reason about the world, which fails to capture the fine-grained physical dynamics and spatial relationships inherent in raw visual data. UniVR addresses this by training a unified generative model to reason natively in visual space, using a reinforcement learning paradigm called VR-GRPO that optimizes for both global task completion and fine-grained physical coherence. On the new VR-X benchmark, UniVR achieves up to a 25% improvement over existing pipelines, demonstrating that models can learn sophisticated policies directly from visual demonstrations without textual mediation.

Paper Primer

Existing visual reasoning systems typically project visual inputs into a linguistic space, forcing the model to rely on text-based chain-of-thought to guide generation. This approach struggles with long-horizon tasks like robotic manipulation or complex physical puzzles, where textual descriptions cannot capture the necessary fine-grained dynamics, leading to logical gaps and physical inconsistencies.

UniVR treats visual reasoning as a next-token prediction task within a unified discrete vocabulary. The core mechanism, Visual Reasoning Group Relative Policy Optimization (VR-GRPO), is a reinforcement learning paradigm: it uses a global reward for task completion and a "Step-Focal" reward that identifies and corrects error-prone substeps by analyzing inter-trajectory variance in latent space. This is like a coach who doesn't just grade the final result, but zooms in on the specific frame where the model's plan diverged from physical reality.

UniVR significantly outperforms state-of-the-art text-guided pipelines on the VR-X benchmark.

UniVR achieves a 60% success rate in long-term planning and 70% in general reasoning, surpassing Gemini 2.5 Pro on key metrics despite having fewer parameters. Up to 25% improvement in overall visual reasoning performance.

Enhanced visual reasoning improves broader multimodal understanding.

Ablation studies show that integrating the UniVR training paradigm leads to consistent gains across standard benchmarks like MMMU and MME, outperforming models trained solely on textual reasoning chains. Significant gains across all six tested multimodal metrics.

Why is a "Step-Focal" reward necessary if a global reward already evaluates task completion?

Global rewards often suffer from "reward hacking," where the model prioritizes terminal success or visual fidelity while ignoring intermediate physical violations or logical gaps in long-horizon sequences.

Does this approach require task-specific heuristics or dense image-text pairs?

No. UniVR learns directly from raw visual demonstration trajectories, bypassing the need for textual supervision or manual task-specific rules.

UniVR demonstrates that reasoning can be effectively offloaded from textual abstractions to native visual-space computation, providing a more grounded and efficient path for training world models.

Introduction and Motivation

We expose the limits of text‑centric reasoning and propose UniVR for native visual‑space reasoning.

Current multimodal models reason in text‑space, which abstracts away the fine‑grained physical dynamics required for complex visual tasks. This abstraction creates logical gaps and physical inconsistencies when models are asked to plan or manipulate in the real world. Consequently, a gap exists between textual reasoning and the demands of native visual understanding.

Reasoning that operates solely on textual symbols, ignoring the visual details of objects, motions, and spatial relations.

We introduce UniVR (Unified Visual Reasoning), the first framework that learns complex reasoning, fine‑grained dynamics, and long‑term planning directly from raw visual demonstrations. At its core lies VR‑GRPO (Visual Reasoning Group Relative Policy Optimization), a reinforcement‑learning paradigm that supplies complementary global and step‑level rewards to enforce logical coherence and physical consistency without any task‑specific heuristics or image‑text pairs.

To evaluate such capabilities we build VR‑X, a large‑scale benchmark curated from 16 diverse sources covering long‑horizon manipulation, spatial puzzles, and general visual reasoning. VR‑X measures both the logical coherence of the generated visual trace and the successful completion of the underlying task.

Across VR‑X, UniVR achieves up to a 25 % improvement over strong baselines, and its visual reasoning also lifts performance on several multimodal understanding benchmarks. These results demonstrate that reasoning directly in visual space can capture dynamics and spatial relationships that text‑centric pipelines miss.

**Figure 1.** UniVR focuses on advancing native visual-space reasoning and planning across diverse tasks. Compared to LMMs that reason within textual space, UniVR facilitates a more profound comprehension of real visual world, improving policy learning in various scenarios.

The key shift is moving from abstract text‑based reasoning to native visual‑space reasoning, which unlocks richer physical and spatial understanding.

The UniVR Framework

UniVR generates visual reasoning traces and trains them with a step‑focused reward.

Purely textual pipelines cannot enforce the fine‑grained physical dynamics required by long‑horizon visual tasks. UniVR replaces that brittle abstraction with native visual reasoning traces that evolve the scene step by step.

A Visual Reasoning Trace is a sequential list of image frames that directly records each intermediate physical state of a task, letting the model “see” how the world changes rather than inferring it from text.

How does a Visual Reasoning Trace differ from a textual chain‑of‑thought?

Chain‑of‑thought strings describe reasoning in language and can omit low‑level dynamics; a Visual Reasoning Trace records every pixel‑level state change, so the model must learn the actual physics rather than just a narrative.

Cold initialization: fine‑tune the unified model on a curated dataset of (query image, instruction, visual reasoning trajectory) pairs.

Reinforcement learning: replace the static supervised loss with the VR‑GRPO objective, letting the model explore and improve its traces.

During RL each sampled trajectory receives the combined reward $R_{\text{reason}}$ and the format reward $R_{\text{format}}$, which are back‑propagated through the policy gradient update.

VR‑GRPO rewards a trajectory like a teacher grading both the final exam (global reward) and each quiz along the way (step‑focal reward), forcing the model to get every intermediate step right.

In what way does VR‑GRPO differ from the original GRPO algorithm?

GRPO only uses a single scalar reward per rollout (typically visual fidelity). VR‑GRPO adds a step‑focal component that explicitly evaluates the most uncertain intermediate segment, and it penalizes divergence between the global and step scores, which forces logical and physical coherence throughout the trace.

t=1: $z_1=(0.1,0.2)$, $z_2=(0.1,0.25)$, $z_3=(0.12,0.22)$ → mean $\bar{z}= (0.107,0.223)$ → variance $\sigma(1)=0.0009$.

t=2: $z_1=(0.5,0.5)$, $z_2=(0.8,0.4)$, $z_3=(0.6,0.7)$ → mean $(0.633,0.533)$ → variance $\sigma(2)=0.045$.

t=3: $z_1=(0.55,0.55)$, $z_2=(0.58,0.52)$, $z_3=(0.57,0.54)$ → mean $(0.567,0.537)$ → variance $\sigma(3)=0.0004$.

t=4 and t=5 yield similarly low variances (<0.001).

Maximum variance occurs at $t^{*}=2$; with window $W=3$ we extract frames $[t=1,2,3]$ as the critical segment for the step‑focal reward.

The step‑focal reward forces the model to resolve precisely the moment where its predictions diverge most, preventing it from “cheating” by getting only the final frame right.

**Figure 2. Overview of UniVR architecture.** Via a unified next-token prediction objective, UniVR processes instructions and image queries to directly generate visual reasoning traces for task execution.

**Figure 3** The proposed Visual Reasoning GRPO. (up) VR-GRPO integrates global and step-focal rewards to ensure both task completion and the physical coherence of generated reasoning traces. (down) Failure case of global rewards that overlook local inconsistencies within long reasoning traces.

The VR-X Benchmark

VR‑X benchmark quantifies visual reasoning across six diverse tasks.

VR‑X comprises 1.5 million raw samples spanning six visual‑reasoning tasks.

The dataset aggregates 1.5 M samples from 16 diverse sources and is curated into 310 k training, 3 k RL, and 1.8 k evaluation samples.

**Figure 1.** Overview of the VR-X framework, demonstrating its capabilities across various domains including robot manipulation, long-term planning, puzzle solving, spatial perception, visual search, and image editing.

**Figure 1.** Overview of the data processing pipeline, consisting of four stages: frame extraction and deduplication, text Q&amp;A generation, frame selection based on relevance, and final quality filtering.

The VR‑X benchmark’s breadth—six heterogeneous tasks—exposes limitations of text‑centric benchmarks and drives progress toward truly visual reasoning.

Experimental Setup

Details of model initialization, training pipeline, and key hyperparameters used in experiments.

We initialize the unified model with the Emu3.5 34B checkpoint and fine‑tune it end‑to‑end via full‑parameter supervised fine‑tuning (SFT) followed by reinforcement learning (RL). The RL pipeline, built on the verl framework, uses a rollout size of 8, a default sub‑step selection window of 4 frames, and a consistency coefficient $\\lambda = 2.0$; training videos are resized to a short side of 512 px and sequences are capped at 20 k tokens. Hyperparameter values (e.g., 62.8, 63.2, 64.1, 67.9, 67.1, 18.2, 22.5, 11.5, 13.5, 28.5, 29.4, 34.7, 42.8, …) are listed in the accompanying table, and Appendix A provides full implementation details.

Performance on VR-X

UniVR outperforms baselines on VR‑X across multiple benchmarks.

Current multimodal models operate in text space, abstracting away fine‑grained physical dynamics. UniVR instead generates visual reasoning traces to preserve spatial accuracy.

UniVR* raises the MathVista score by 18.2 points over Emu3.5.

Table 2 shows MathVista 799.3 for UniVR* versus 781.1 for Emu3.5.

Emu 3.5 is a text‑only trained vision‑language model that relies on textual instructions to drive image generation.

**Table.** Comparison of Emu 3.5, Text-only training, and UniVR* across various benchmarks including MMMU, MME(P), MME(C), MMBench, MathVista, and MM-Vet.

**Figure 7.** Comparison with Gemini and Emu3.5. In each group, the first, second, and third rows correspond to Gemini 3 Pro + Nano Banana 2, Emu3.5, and UniVR, respectively.

Overall, UniVR demonstrates that training directly in visual space yields substantial gains, confirming the paper’s premise that visual‑space reasoning is essential for complex multimodal tasks.

Ablation Studies

Ablations isolate each component’s impact on multimodal performance.

This section quantifies how each design choice contributes to the overall gains reported earlier. We report results from Tables 2a–2e, each ablating a single component while keeping the rest fixed.

Applying the global reward alone improves short‑sequence reasoning by 3.6 % but degrades long‑term planning and JEPA scores, indicating reward hacking.

Table 2b, row 2 (global reward only)

UniVR raises baseline performance on three external visual‑reasoning benchmarks by up to 24.3 %.

Table 2e (external benchmark evaluation)

**Figure 6. Analysis of the Visual Reasoning Reward.** (Up) Relying solely on the global reward misses step-level errors in long-horizon planning tasks, e.g. hangers penetrating garments (top row), erroneous liquid pouring dynamics (middle row), and jittering transitions between paper towels (bottom row). (Down) RL reward curves with and without the step-level reward component.

Appendices

Appendix details the data pipeline, training stages, and evaluation challenges underlying UniVR.

Current multimodal systems train on text‑centric objectives, which discard the fine‑grained physical dynamics needed for complex visual tasks; as a result, generated sequences often contain logical gaps, implausible interactions, and error accumulation over long horizons.

The appendix reveals three concrete obstacles: (1) building a high‑quality visual‑reasoning corpus requires a multi‑stage pipeline that discards ~80 % of raw video candidates; (2) training on 15 k‑token trajectories for 34 B‑scale models consumes massive GPU resources; and (3) naïve global rewards encourage shortcut‑seeking behavior, whereas the step‑level VR‑GRPO reward is needed to keep intermediate reasoning coherent.

Read the original paper

Open the simplified reader on Paperglide

Browse all simplified papers