PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2607.12800v1cs.CV

TL;DR

UniVR addresses the limited ability of existing approaches to learn complex reasoning, physical dynamics, and long-term planning directly from visual data. It introduces visual-space modeling, VR-GRPO, and the VR-X benchmark, achieving stronger performance across diverse visual reasoning tasks, including 60% success in long-term planning and 70% in general reasoning.

  • Problem

    Existing AI systems primarily reason and plan in text, which cannot fully represent visual-world dynamics, spatial relationships, and physical laws.

  • Method

    UniVR models visual state transitions from image-sequence demonstrations and trains them with VR-GRPO’s global and step-level rewards on the diverse VR-X benchmark.

  • Results

    UniVR enhances performance across VR-X tasks, reaching 60% success in long-term planning and 70% in general reasoning.

  • Takeaways & Limitations

    Raw visual demonstrations can support sophisticated long-horizon policy learning, while stronger visual reasoning also benefits multimodal understanding.

Abstract

from arXiv · show

Learning broad world knowledge directly from raw visual data is a fundamental capability of intelligence. We introduce UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and long-term planning from pure visual demonstrations. At its core, UniVR features VR-GRPO, a reinforcement learning paradigm with complementary global and step-level rewards. This approach enforces logical coherence and physical consistency throughout the reasoning process without requiring task-specific heuristics or image-text pairs. To train and evaluate UniVR, we construct VR-X, a large-scale benchmark curated from 16 diverse sources spanning long-horizon manipulation, spatial puzzles, and physical reasoning. It is the first comprehensive suite to assess these heterogeneous capabilities under a purely visual protocol. Remarkably, UniVR achieves up to a 25% improvement on VR-X, and its superior visual reasoning also boosts performance on various multimodal understanding benchmarks. These findings underscore the vast potential of reasoning within visual spaces, with all code, data, and models are open-sourced for further research.

1 Introduction

UniVR addresses the limits of text-mediated visual reasoning by learning heterogeneous reasoning and planning directly in visual space. It introduces VR-X and VR-GRPO to improve logical coherence and physical consistency across diverse visual tasks.

  • Motivation: Text-based reasoning struggles to represent complex dynamics, spatial relationships, and physical laws, causing failures in long-horizon visual evolution.Existing systems often produce logical gaps and physical inconsistencies despite strong textual reasoning or visual generation capabilities.
  • Benchmark: VR-X evaluates visual reasoning across long-horizon planning and general reasoning tasks, including fine-grained dynamics in cooking, crafts, robotics, and navigation.Its planning tasks include minute-scale activities such as tying knots and folding clothes.
  • UniVR: UniVR learns heterogeneous tasks directly in a unified visual space without language supervision.The framework targets capabilities ranging from long-term planning to general cognitive reasoning.
  • VR-GRPO: VR-GRPO combines a global reward with Step-Focal rewards that target error-prone reasoning substeps.This design improves logical coherence and physical consistency without image-text pairs or task-specific rules.
  • Results: With only 34B parameters, UniVR approaches the Gemini 3 Pro plus Nano Banana 2 pipeline and surpasses Gemini 3 on long-horizon manipulation.The reported comparison indicates improved visual reasoning while preserving the base model’s foundational strengths.

2 Related Works

Prior visual reasoning methods extend text-based reasoning to multimodal inputs but remain bounded by linguistic representations, while unified models seek a shared space yet are constrained by training objectives. Reinforcement learning has improved visual generation and multimodal understanding, but existing rewards mainly target perceptual quality, semantic alignment, or single-step correctness.

  • Visual Reasoning: Visual CoT extends chain-of-thought reasoning to multimodal inputs, but projects visual features into linguistic space and remains fundamentally text-bounded.This work is positioned against text-centric and latent-space reasoning paradigms that motivate visual reasoning extensions.
  • Unified Model: Unified models combine textual reasoning with visual generation in a shared model space, yet their visual knowledge acquisition is constrained by entertainment-oriented or artistic editing objectives.Pioneering architectures tokenize images, text, and video into a unified discrete space for cross-modal generation.
  • Reinforcement Learning in Generative Model: Reinforcement learning methods such as DDPO and ReFL align visual generation with human preferences for image fidelity using PPO or RLHF.These approaches extend reinforcement learning in multimodal language models to visual generation.
  • Reinforcement Learning in Generative Model: Existing GRPO-based vision methods use specialized rewards for aesthetic appeal, semantic alignment, or single-step correctness, unlike rewards designed for broader reasoning consistency.Examples include HPSv3 for aesthetics and CLIP- or VLM-based scoring for semantic alignment.

3 UniVR

UniVR uses autoregressive visual-space generation trained through cold initialization and reinforcement learning on demonstration trajectories. Its VR-GRPO method combines global and step-focal rewards to enforce task completion, logical coherence, and physical consistency.

  • Training framework: UniVR follows a two-stage training pipeline consisting of cold initialization followed by reinforcement learning.The framework uses autoregressive generation as its basic model architecture.
  • Visual-space reasoning: UniVR models next-frame distributions from image-sequence demonstrations, directly learning state transitions and policy dynamics without dense textual reasoning chains.This formulation encourages reasoning within visual space across diverse planning and reasoning scenarios.
  • Visual Reasoning GRPO: VR-GRPO addresses heterogeneous multi-source tasks by avoiding dense image-text pairs and task-specific heuristics while targeting logical coherence and task completion.Vanilla SFT struggles to reconcile differences in temporal scale, domain knowledge, and visual appearance across tasks.
  • Reward design: VR-GRPO combines format, global, and step-focal visual reasoning rewards to enforce structural constraints, task completion, and physical coherence.The step-focal reward targets error-prone trajectory steps, complementing global assessment that can overlook intermediate violations and logical gaps.
  • Reward design: The step-focal reward identifies peak uncertainty in reasoning trajectories and focuses evaluation on critical segments to prevent shortcuts and preserve procedural integrity.Per-frame CLIP embeddings estimate inter-trajectory variance, with high variance indicating divergent reasoning paths.

4 VR-X Benchmark

VR-X is a large-scale benchmark for heterogeneous visual reasoning, covering six task types with detailed reasoning traces. It spans diverse sources and time horizons, while evaluating both logical accuracy and physical consistency.

  • Benchmark scope: VR-X covers six visual reasoning tasks: guidance, robotic manipulation, puzzles, editing, search, and spatial perception.Each task provides detailed visual reasoning traces.
  • Data composition: 1.5M raw samples from 16 diverse sources span minute-long planning and single-step reasoning.The sources include AgiBot, Action100M, EgoDex, and VisualCoT.
  • Data composition: 310k cold-start training, 3k RL, and 1.8k benchmark evaluation samples are rigorously curated from the raw data.All samples follow a unified visual reasoning trajectory.
  • Data composition: Fine-grained textual chain-of-thought descriptions annotate the visual reasoning sequences to support multimodal learning.The annotations accompany the unified visual reasoning trajectories.
  • Evaluation: Evaluation measures logical accuracy and adherence to real physical dynamics using VLM score and JEPA similarity.VLM score aggregates task completion, procedural coherence, visual informativeness, and image fidelity into a normalized 0–100 score; JEPA similarity addresses fine-grained physical inconsistencies.

5 Experiment

Experiments on VR-X compare UniVR with multimodal and unified-generation approaches, showing strong visual-space reasoning performance. Ablations further examine VR-GRPO’s reward components and UniVR’s effects on multimodal understanding, heterogeneous-task training, and external benchmark generalization.

  • VR-X comparison: Most unified generation models reach at most 30% success because they lack contiguous image generation and iteratively unroll long-horizon plans.Emu3.5 performs best among these models but stagnates at 35% on VR-X’s intricate visual dynamics.
  • VR-X comparison: UniVR achieves 60% success in long-term planning and 70% in general reasoning without fine-grained text procedural annotations.At 34B parameters, it outperforms Gemini 2.5 Pro on several key metrics.
  • Additional evaluations: UniVR’s visual reasoning training is evaluated for effects on multimodal understanding, heterogeneous-task balance, and generalization to three external visual reasoning benchmarks.The external evaluation covers two embodied-intelligence benchmarks and one cognitive-reasoning benchmark using the same VLM scoring protocol.
  • VR-GRPO ablation: Global reward alone improves general reasoning by 3.6% but degrades long-term planning and JEPA scores, whereas step-focal reward improves all three metrics.Combining both rewards has a synergistic effect, and pairwise VLM scoring further improves reward robustness and final results.
  • Reward compatibility: Textual reasoning chains, HPSv3, and CLIP rewards provide limited assistance for complex visual reasoning, while VR-GRPO is compatible with text-based reinforcement learning.The passage reports no significant gains from adding textual reasoning chains during cold start and negligible impact from HPSv3.

6 Conclusion · Appendix

UniVR enables complex visual-space reasoning and planning without dense language supervision, using VR-GRPO’s complementary global and step-level rewards. The appendix adds implementation details, benchmark and metric analysis, and further ablations with visualizations.

  • 6 Conclusion: UniVR enables complex reasoning and planning directly within visual space, without dense language supervision.
  • 6 Conclusion: VR-GRPO combines global task-completion rewards with fine-grained step-level physical-coherence rewards.
  • 6 Conclusion: UniVR is trained and evaluated on the diverse VR-X benchmark.
  • 6 Conclusion: UniVR achieves substantial gains over strong text-based reasoning pipelines and unified generation models.
  • Appendix: Appendix A covers implementation details of the two-stage training pipeline.
  • Appendix: Appendix B presents further analysis of the VR-X benchmark and evaluation metrics.
  • Appendix: Appendix C includes additional ablation studies and visualizations.

A Implementation Details · A.1 Cold Initialization

UniVR’s cold initialization uses a multi-stage pipeline to curate visual training data, then trains from Emu3.5-34B with full-parameter optimization and bounded visual-token sequences. The process yields 310k curated samples from 1.5M raw candidates while accommodating multi-minute reasoning trajectories.

  • A.1 Cold Initialization: Training data is constructed from multiple sources whose subsets are weighted according to sample counts.The data-construction pipeline filters raw data into curated training samples through four stages.
  • A.1 Cold Initialization: Raw video sequences undergo scene-aware temporal sampling with PySceneDetect at 0.27 FPS before curation.This is the first stage of the four-stage data-processing pipeline.
  • A.1 Cold Initialization: Non-video VisualCoT and ZebraCoT image sequences bypass temporal sampling and are standardized to match video-derived data.These sources are already formatted as image sequences.
  • A.1 Cold Initialization: 310k cold initialization samples are curated from 1.5M raw candidates.The curated set combines data processed from the multiple source types described in the pipeline.
  • A.1 Cold Initialization: Cold initialization starts from Emu3.5-34B and trains UniVR with full parameters on 32 GPUs.The optimization uses the configurations specified in Tab. 3’s second column.
  • A.1 Cold Initialization: Images are resized to 512 on the short side and tokenized with Emu3.5’s VQ tokenizer into 1,000–1,500 tokens per image.These preprocessing choices define the visual input representation during cold initialization.
  • A.1 Cold Initialization: The maximum sequence length is capped at 15,000 to reduce training overhead while accommodating trajectories spanning approximately several minutes.The cap is applied per sample.

A.2 Reinforcement Learning

UniVR’s reinforcement-learning stage filters challenging visual-reasoning samples and trains from an SFT checkpoint with full-parameter optimization. The setup uses multi-GPU training with Qwen3-VL-30B-A3B evaluation and disables CFG during rollout.

  • Data Construction: Approximately 3k hard samples are filtered from the original 310k training set using the post-initialization model.The subset contains roughly 2k long-term planning trajectories and 1k general reasoning samples.
  • Data Construction: Roughly 2k long-term planning trajectories predominantly span 6–10 reasoning steps, while 1k general reasoning samples preserve task diversity.These samples are selected to support complex reasoning procedures in visual reasoning reinforcement learning.
  • Optimization: Reinforcement learning initializes from the SFT checkpoint and trains full parameters on 32 GPUs using the configuration in Tab. 3.A Qwen evaluator runs on an additional 8 GPUs.
  • Optimization: Qwen3-VL-30B-A3B serves as the evaluator to improve training efficiency, while larger variants provide more accurate rewards but significantly increase training latency.Classifier-Free Guidance is disabled during rollout.

B Details on VR-X Benchmark

VR-X evaluates visual reasoning using 1.8k professionally annotated reasoning trajectories and two complementary metrics: VLM score and JEPA similarity score. The VLM protocol assesses multiple reasoning dimensions and aligns well with human judgments, while JEPA measures distributional similarity between visual sequences.

  • VR-X Evaluation Set: 1.8k professionally annotated reasoning trajectories comprise the held-out VR-X evaluation set, which uses VLM and JEPA similarity scores.The trajectories are curated from a held-out subset, with detailed composition reported in Table 4.
  • VLM Evaluation: The VLM score evaluates visual quality, task completion, logical coherence, physical dynamics, and temporal consistency using both ground-truth and generated sequences.A unified prompt lets the VLM reference the ground-truth logic process and action dynamics.
  • VLM Evaluation: Approximately 0.85 correlation between human and VLM scores indicates high alignment with professional annotator judgments.Professional annotators first score model outputs, after which the samples are shuffled and presented for VLM evaluation.
  • JEPA Evaluation: The JEPA score replaces I3D features with a V-JEPA encoder and computes polynomial-kernel MMD between 1280-dimensional latent-vector distributions, where smaller distances indicate higher similarity.This computation resembles traditional video metrics such as FVD.

C More Analysis and Visualizations

Additional analyses show that VR-GRPO mitigates reward hacking and stabilizes long-horizon visual reasoning by preserving intermediate logical and physical consistency. Comparisons and visualizations highlight UniVR’s stronger execution coherence, while limitations remain in computational cost and reward-model physical knowledge.

  • Reward hacking: Global reasoning rewards can overlook localized physical and logical errors in 30+ second sequences, whereas visual reasoning rewards expose these failures.Examples include implausible hanger-garment interactions, incorrect wine-pouring dynamics, and flawed towel-replacement logic.
  • Long-horizon stability: VR-GRPO’s step-level rewards deliver the most pronounced gains on >30s sequences by correcting error-prone intermediate steps.The evaluation partitions samples into <10s, 10–30s, 30–60s, and >60s duration groups.
  • Long-horizon stability: Step-level rewards gate compounding errors during long-range prediction, producing superior stability on extended reasoning traces.This mechanism explicitly maintains logical coherence and physical consistency at intermediate stages.
  • Baseline comparisons: Compared with Gemini 3 Pro + Nano Banana 2 and Emu3.5, UniVR preserves comparable visual fidelity while improving logical coherence and physical consistency.Baselines show execution inaccuracies in rope manipulation, garment unfolding and folding, paper-folding dynamics, and environmental consistency.
  • Additional visualizations: UniVR generates long-horizon sequences with coherent logic, physical dynamics, and temporal consistency across additional scenarios.These results are presented in Fig. 8.
  • Limitations: Training 34B-scale models on long visual sequences requires substantial computation, while the reward mechanism relies on general-purpose VLMs with limited fine-grained physical-world knowledge.The authors identify a more powerful reward system natively grounded in visual-world dynamics as a needed direction.
Loading 2607.12800v1…