PaperlayerView exact revision on arXiv

Source-linked AI summary

VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

Junxiang Xu, Ruisi Wang, Fanyi Pu, Maijunxian Wang, Ran Ji, Tongxi Zhou, Chenyang Gu, Jing Zuo, Hongcan Xiao, Yimeng Geng, Wanqi Yin, Wei Chen, Oscar Qian, Zhengan Yan, Ziqi Huang, Haiwen Diao, Liang Pan, Bo Li, Xiangyu Fan, Dezhi Luo, Fengyuan Yu, Zehong Zhao, Qingying Gao, Tinghui Zhu, Yilan Zhang, Jingqi Tong, Pinyuan Feng, Zhengze Jiang, Letian Wang, Ziyu Guo, Renrui Zhang, Jieneng Chen, Sonia Joseph, Constantin Venhoff, Saman Motamed, Mengyue Yang, Chandra Sripada, Alan Yuille, Philip Torr, Lvmin Zhang, Vikash Kumar, Daniel Khashabi, Nikolaus Kriegeskorte, Raphaël Millière, Vincent C. Müller, Anyi Rao, Quan Wang, Ziwei Liu, Dahua Lin, Lei Yang, Hokin Deng, Zhongang Cai

arXiv:2608.26105v1cs.CVcs.AIcs.LGcs.MMcs.RO

TL;DR

Native visual reasoning lacks scalable training tasks, reliable feedback, and controlled comparisons across visual generative substrates. VBVR-Pro addresses these gaps with a procedurally generated task suite, deterministic verifiable rewards, and controlled studies of image, video, and interleaved generation. The resulting analyses show transfer beyond the suite, stronger video performance for persistent spatiotemporal tracking, and a compute-efficient role for interleaved generation.

  • Problem

    Native visual reasoning remains bottlenecked by limited scalable training tasks, reliable feedback, and controlled comparisons across generative substrates.

  • Method

    VBVR-Pro unifies procedurally generated visual reasoning environments, task-grounded verifiable scorers, reinforcement learning, and modality-controlled analyses.

  • Results

    Models trained on VBVR-Pro transfer beyond the suite, while video is strongest for persistent spatiotemporal tracking and interleaved generation offers a compute-efficient alternative.

  • Takeaways & Limitations

    VBVR-Pro makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable.

  • Takeaways & Limitations

    VLM-as-a-judge remains unreliable for visual reasoning evaluation and reinforcement-learning rewards because of inaccurate and non-reproducible judgments.

Abstract

from arXiv · show

Native visual reasoning treats visual generation as the medium of reasoning itself: visual states (i.e. images and videos) are not merely inputs to be understood or outputs to be rendered, but first-class substrates for problem solving beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across seven external visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded in deterministic, task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative. Critically, ablations and probing suggest the presence of vision-native trajectories that are crucial to visual reasoning. We release all data, models, scorers, and code.

1. Introduction

VBVR-Pro addresses the lack of scalable tasks, reliable feedback, and controlled comparisons that constrain native visual reasoning research. It provides a procedurally generated curriculum, verifiable rewards, and a modality-controlled testbed for studying visual states as reasoning substrates.

  • VBVR-Pro creates a controlled task space of 300 procedurally generated visual reasoning tasks with aligned image, video, and interleaved modalities.The shared task distribution supports fair comparison among different generators and records metadata for verifiable evaluation.
  • Models trained on VBVR-Pro obtain consistent downstream gains, often exceeding 20 percentage points across external visual reasoning benchmarks.
  • Task diagnostics suggest VBVR-Pro functions as a scalable curriculum rather than merely an enlarged task collection.
  • Task-grounded verifiable scorers address recurring VLM-judge failures involving numerical imprecision, fine-grained evidence, and misunderstood rules.The scorers use structured semantic extraction and task-grounded verification, with instance-level human evaluation showing strong agreement and higher per-instance accuracy than VLM judges.
  • Controlled studies across more than 30 generators find video strongest for persistent spatiotemporal tracking, while interleaved generation is more compute-efficient.Image-only generation often lacks capacity for procedural or temporal reasoning, and visual-state interventions support the use of intermediate visual states.
  • VBVR-Pro combines scalable curricula, human-aligned rewards, reinforcement learning, and modality-controlled analyses into a closed-loop foundation for native visual reasoning.

2. VBVR-Pro-Dataset: Task Scaling for Transferability

VBVR-Pro-Dataset targets limited transfer from narrow synthetic task sources by expanding task coverage across diverse visual environments and reasoning capabilities. Its procedural generators produce aligned modalities, verifiable metadata, and curated training and evaluation splits.

  • Task Scaling for Transferability: VBVR-Pro-Dataset contains 300 tasks spanning diverse visual environments and reasoning capabilities to improve transfer beyond training patterns.The design treats task coverage, rather than instance count alone, as a central bottleneck in scalable visual reasoning training.
  • Task Design: The taxonomy organizes tasks into five faculties: perception, spatiality, transformation, abstraction, and knowledge.
  • Task Design: Tasks are designed to depend on vision-native transformations, spatial relations, or dynamics and to produce unambiguous, inspectable outcomes with visual diversity.
  • Implementation: 300 task-specific generators sample structured parameters, instantiate problem instances, and use task-specific solvers to provide programmatically derived supervision.Half of the generators are reimplemented from VBVR and half are newly designed to broaden coverage.
  • Implementation: Aligned metadata stores seeds, specifications, solutions, and element attributes for verifiable scoring, deduplication, and train/evaluation split construction.
  • Implementation: Image outputs use Last-Frame, Key-Frame, or Multi-Frame regimes according to whether final state, essential transitions, or the full solution process must be evaluated.Videos and images are different renderings of the same underlying problem with equivalent solutions, enabling controlled modality comparisons.
  • Statistics and Analysis: VBVR-Pro holds out 50 tasks for out-of-domain evaluation and samples 5,000 instances from each of 250 remaining tasks, yielding 1.25M training instances.Its benchmark includes 50 out-of-domain and 50 in-domain tasks, totaling 100 tasks.
  • Statistics and Analysis: 47% of newly developed tasks require multi-step reasoning, compared with 7% of reworked VBVR tasks, supporting broader and more challenging coverage.

3. Verifiable Reward Scorers

VBVR-Pro replaces unreliable VLM judges with deterministic, task-specific verifiable reward scorers designed for precise, reproducible evaluation and reinforcement-learning feedback. The scorers use semantic checks and achieve stronger human alignment, lower evaluation cost, and high model-ranking correlation.

  • Limitations of VLM-as-a-Judge: VLM judges can overestimate incorrect outputs, overlook decisive visual errors, and underestimate correct answers across fine-grained reasoning tasks.Observed failures include incorrect hue, changed preserved circle attributes, and inconsistent scoring of a correct pattern-completion answer.
  • Limitations of VLM-as-a-Judge: VLM judges are costly and difficult to use at reinforcement-learning scale, whereas task-specific scorers are lightweight and substantially cheaper per evaluation.The cost burden comes from proprietary API calls or open-source GPU inference competing with model training.
  • Limitations of VLM-as-a-Judge: 55–93% of repeated VLM evaluations change scores at temperature 0, while the deterministic scorer produces identical results across runs.This matters for per-sample rewards because inconsistent feedback can create noisy or conflicting optimization signals.
  • Design of Verifiable Reward Scorers: Each of the 100 benchmark tasks receives a dedicated scorer that extracts semantic attributes and applies calibrated task-specific checks.Implemented checks include color segmentation, contour detection, OCR, and trajectory tracking; soft criteria use weighted sums, while hard constraints use multiplicative aggregation.
  • Validation Against Human Preferences: Per-vote agreement exceeds 0.60, surpassing GPT-5.5 at 0.54 and Gemini-3.1-Pro at 0.52, while reaching approximately 78% of the human agreement ceiling of 0.77.The scorer also has the lowest evaluation cost among the compared evaluators.
  • Validation Against Human Preferences: The scorer reaches Spearman’s ρ = 0.95 in-domain and ρ = 1.00 out-of-domain, with Pearson’s r ≥0.95 on both splits when ranking generation models.Out-of-domain correlation improves over the earlier VBVR scorer from ρ = 0.93 to 1.00 and r = 0.92 to 0.95.

4. Benchmarking Native Visual Reasoning

VBVR-Pro-Bench evaluates visual reasoning across generative paradigms, task distributions, and interventions. Training improves within-task generalization, cross-task transfer, and evidence for reusable visual-state reasoning, while video is especially useful for state transitions.

  • Benchmark scope: More than 30 models are benchmarked across image, video, and interleaved generation, with diagnostics and transfer evaluations beyond the suite.The evaluation includes overall performance, counterfactual interventions, Chain-of-Step analysis, and seven external reasoning benchmarks.
  • Main results: 0.290 average overall gain follows task-specific training across all nine evaluated models and three generation modalities.The strongest trained model nevertheless remains well below human-level performance.
  • Generalization: +0.401 in-domain and +0.179 out-of-domain average gains show improvements in both within-task generalization and transfer across task families.Out-of-domain evaluation uses different task distributions from training.
  • Generative paradigms: The strongest interleaved model performs on par with the strongest video model across many in-domain tasks after task-specific training.This indicates that sparse, task-relevant visual states can match video performance when the underlying task structure is learned.
  • Generative paradigms: Video models outperform other paradigms when reasoning requires accurate state transitions, including transformation tasks and illustrated multi-step manipulations.They generate coherent intermediate states and reach correct solutions where interleaved and single-image models fail.
  • Visual trajectories: 0.024 and 0.111 overall-score reductions occur when visual trajectories are compressed into single images.These ablations retain the original reasoning text, isolating the contribution of intermediate visual supervision.
  • Visual trajectories: Intermediate visual states are more consequential than intermediate text during inference, supporting visual-state maintenance and updating as the working representation.Removing or corrupting intermediate images causes larger drops than removing or contradicting intermediate language.
  • Transfer: External evaluation shows improvements of up to 20 percentage points across seven unseen benchmarks, including real-world and embodied scenarios.These benchmarks differ substantially from the procedurally generated abstractions in VBVR-Pro-Dataset.

5. Reinforcement Learning with Verifiable Rewards

VBVR-Pro studies reinforcement learning for native visual reasoning under a controlled setup, focusing on semantic exploration and task-grounded rewards. CPS preserves rollout fidelity while enabling exploration, and RLVR outperforms VLM-judged RL and SFT with more sustained gains and improved out-of-domain performance.

  • Motivation: Native visual reasoning requires RL rewards for semantic transformations and task-specific constraints rather than pixel-level fidelity or aesthetics.The RL baseline is designed for high-level task goals expressed through visual transformations.
  • Exploration: Standard stochastic perturbations often fail to explore reward-relevant decisions, while excessive stochasticity damages visual fidelity.Useful exploration includes discrete choices such as maze turns, answer selection, and object connections.
  • Exploration: CPS preserves total scheduler noise coefficients while replacing part of predicted noise with Gaussian noise, balancing semantic exploration and visual quality.CPS uses η to control stochasticity and reduces to deterministic ODE sampling at η = 0.
  • Experimental design: RLVLM and RLVR differ only in reward source under the same algorithm, rollout configuration, hyperparameters, data, and initialization.This isolates the effect of replacing a general-purpose VLM judge with the task-grounded VBVR Scorer.
  • Results: 0.548 overall score for RLVR exceeds 0.508 for RLVLM and 0.503 for SFT under matched CPS sampling with η = 0.7.RLVR produces the strongest performance in the controlled comparison.
  • Results: 0.377 out-of-domain score for RLVR exceeds 0.345 for RLVLM and 0.328 for SFT, despite no out-of-domain tasks during RL training.The results indicate improved generalization beyond the RL training distribution.
  • Mechanism: RLVR extends effective reasoning within a fixed denoising budget and improves temporal consistency of multi-step visual trajectories.Qualitative examples show revisable color decisions and connected maze routes leading to the target.

6. Related Work

Related work spans native visual generation, interleaved visual-textual reasoning, reinforcement learning, and visual reasoning benchmarks, while highlighting gaps in training coverage and verifiable evaluation.

  • Native visual reasoning through visual generation: Native visual reasoning uses generated visual states as reasoning substrates, including video diffusion trajectories and image-to-image generation.Prior work studies Chain-of-Frames, Chain-of-Steps, and native image-generation approaches for visual reasoning.
  • Interleaved visual-textual reasoning: Interleaved reasoning methods combine intermediate visual states with textual rationales through sketches, visualized traces, editing, or interleaved corpora.
  • Reinforcement learning for visual reasoning: Reinforcement learning has mainly targeted instruction following and aesthetics, leaving semantically valid visual reasoning objectives comparatively underexplored.
  • Visual reasoning benchmarks and training sources: Existing benchmarks cover video, image editing, and unified multimodal generation but are mostly evaluation-only, weakly supervised, difficult to verify, or limited in procedural task coverage.Some interleaved corpora reach training scale, but they are not procedurally generated and remain limited in task coverage.

7. Conclusion

VBVR-Pro unifies procedural environments, verifiable rewards, reinforcement learning, and controlled modality analysis into a trainable and experimentally controllable framework for native visual reasoning. Its model suite spans image, video, and interleaved generators with diverse architectures and representations.

  • 7. Conclusion: VBVR-Pro combines procedural visual reasoning environments, task-grounded rewards, reinforcement learning, and modality-controlled mechanism analyses.The framework is presented as trainable, verifiable, optimizable, and experimentally controllable.
  • 7. Conclusion: The study suggests that visual generative models can construct, maintain, and refine intermediate visual trajectories rather than treating generation only as an output format.
  • 7. Conclusion: The evaluated generators span image, video, and interleaved text-image models with substantial variation in scale, cross-modal interaction, and output representation.
  • Image generation models: Qwen-Image-Edit and FLUX.2-dev use multimodal diffusion Transformers, while FLUX.2-dev transitions from separate processing to joint single-stream processing.

A.1.1. Training Recipes

Training uses modality- and scale-dependent fine-tuning recipes, with full fine-tuning for smaller image or interleaved models and LoRA for larger or video models.

  • Training Recipes: All models are fine-tuned for one epoch at 512 × 512 resolution, using full fine-tuning below 10B active parameters and LoRA for larger image, interleaved, and all video models.
  • Training Recipes: Qwen-Image-Edit and FLUX.2-dev use rank-32 LoRA at 1 × 10^-4, whereas BAGEL is fully fine-tuned at 5 × 10^-6.
  • Training Recipes: VBVR-SenseNova-U1 and VBVR-Pro-SenseNova-U1 start from an unreleased checkpoint never exposed to VBVR data, isolating gains attributable to fine-tuning.

B.1. Training Pipeline

The training pipeline overlaps CPU reward computation with GPU rollouts and optimization through a one-step delay, reducing idle time while bounding policy staleness.

  • Training Pipeline: The synchronous pipeline leaves GPUs idle while rollout preprocessing and reward evaluation finish before policy optimization.
  • Training Pipeline: A fixed one-step delay asynchronously computes rewards for rollout k while GPUs generate rollout k + 1, then optimizes using rollout k.
  • Training Pipeline: The delayed design preserves deterministic execution order and limits policy staleness to at most one rollout iteration.
  • Training Pipeline: Generating samples at 512 × 512 and upscaling only during reward computation reduces trajectory-generation cost while retaining high-resolution reward inputs.
  • Training Pipeline: CPU-based verifiable scoring overlaps with GPU work, whereas GPU-resident RLVLM judging reserves 50% of available GPU memory.
  • Training Pipeline: KL regularization is omitted because it does not provide consistent gains in this setting.

B.3. Training and Inference Configuration

The experiments use matched and explicitly varied inference configurations to compare reinforcement-learning methods and sampling solvers. RL evaluations generally use 30 denoising steps, while the SFT baseline retains Wan’s default UniPC setup.

  • Training configuration: RL training uses CPS with 30 sampling steps, η = 0.7, 60% transition replay, batches of 32 prompts, and 32 trajectories per prompt.The learning rate is 5 × 10−6 and CFG scale is 1.0 unless otherwise specified.
  • Matched comparison: Verifiable-RL and VLM-Judge RL use the same CPS configuration with η = 0.7, isolating reward-objective effects from sampler differences.Both evaluations use 30 denoising steps, matching training rollouts.
  • Inference configurations: The sampler-sensitivity study compares CPS against deterministic ODE sampling with Euler and UniPC under the same 30-step schedule and CFG scale.This keeps the solver as the only changed inference variable.
  • Baseline: The SFT baseline retains Wan’s default UniPC inference configuration with 50 sampling steps and CFG scale 5.This configuration is unchanged from the original model inference pipeline.

B.4. Sampler Sensitivity

Sampler sensitivity shows that moderate CPS stochasticity benefits Verifiable-RL, while higher stochasticity can destabilize VLM-Judge RL. The study compares these samplers across overall, in-domain, and out-of-domain performance.

  • Study design: CPS η = 0.7 is the matched main-paper configuration for both reinforcement-learning methods, while the extended study includes η ∈{0.1, 0.3, 0.7, 0.9} and two ODE solvers.The comparison reports overall, In-Domain, and Out-of-Domain scores.
  • Verifiable-RL: 0.548 overall score is achieved by Verifiable-RL with CPS η = 0.7, up from 0.509 at η = 0.1.Both In-Domain and Out-of-Domain scores also improve, while performance decreases slightly at η = 0.9.
  • Pipeline: The one-step-delayed pipeline overlaps asynchronous preprocessing and reward evaluation with rollout generation and policy optimization, reducing GPU idle time while bounding policy lag to one rollout iteration.
  • VLM-Judge RL: VLM-Judge RL shows smaller and less consistent sampler differences, with higher-stochasticity CPS reaching a higher ceiling but suffering a temporary drop around Steps 400–500.Deterministic ODE trajectories are smoother.

B.5. Analysis of VLM-Judge Instability

VLM-Judge RL’s temporary instability under stochastic CPS is associated with background-color degradation rather than loss of foreground reasoning. Qualitative trajectory analyses show Verifiable-RL revising intermediate predictions and producing more coherent temporal behavior.

  • Sampler-dependent degradation: At the unstable checkpoint, foreground transformations remain largely correct while increasingly gray backgrounds appear as CPS stochasticity rises.Low-stochasticity CPS and deterministic ODE sampling better preserve the target white background.
  • Failure evolution: Background-color deviation emerges at Step 400, peaks at Step 500, and largely recovers by Step 600 under CPS η = 0.9.This timing closely matches the temporary evaluation-performance drop.
  • Reward reliability: The general-purpose VLM judge may fail to penalize subtle background deviations, allowing them to persist during optimization and amplify under stochastic sampling.The task-grounded scorer explicitly captures these deviations and provides a more stable optimization signal.
  • Trajectory visualization: The visualization recovers agent trajectories by extracting connected-component centroids, interpolating isolated failures, and preserving longer disappearances or large jumps.Trajectories are divided at task-specific targets and rendered as colored directed stages over a median background using all 81 frames.
  • Intermediate revision: Verifiable-RL redirects an initially selected triangle toward the correct blue pentagon, unlike the pre-RL model, under the same 30-step schedule.The comparison attributes the difference to denoising-trajectory use rather than additional inference computation.
  • Trajectory coherence: Across set-intersection, object-matching, and waypoint-navigation examples, Verifiable-RL progressively refines predictions and produces more continuous multi-stage behavior.The pre-RL model remains focused on irrelevant boundaries, inconsistent pairings, or incoherent intermediate goals.

C.2. Interleaved Data Generation

Interleaved data pairs each generated image with prospective textual reasoning, while the broader dataset analysis evaluates task complexity and reasoning depth across reworked and newly designed tasks. The new tasks are reported as visually richer and more reasoning-intensive.

  • Interleaved representation: Each instance containing K output images produces K reasoning-image pairs, with each text step placed immediately before its corresponding image.The images come directly from the image modality.
  • Reasoning generation: Gemini-3.1-Pro-Preview generates one prospective reasoning step per solver-provided output image using the complete solution trajectory.The model annotates known state transitions rather than independently discovering the solution.
  • Quality control: Task-specific prompts specify rules, image order, required reasoning, and output format, with five annotations per task inspected for consistency with solver transitions.The checks ensure each step explains the decision leading to its output image rather than merely captioning it.
  • Visual complexity: 6.9× as many distinct regions and 2.4× as many distinct colours characterize newly designed tasks versus reworked tasks by median task-level values.New tasks also show 1.7× higher spatial information and a 2.0× larger encoded size; the conclusion relies on all four measures.
  • Output regimes: The three image output regimes are Last-Frame, Key-Frame, and Multi-Frame, with path-based tasks representing complete trajectories as overlays on a selected frame.The regime depends on how much of the solution trajectory must be explicit.
  • Reasoning depth: 75.3% of all task pairs selected the newly designed task as requiring more reasoning under both presentation orderings, versus 13.3% selecting the reworked task.The orderings agreed for 133 of 150 pairs, or 88.7%.
  • Human evaluation: Human preference annotations rate each video independently on four levels: L0 fully correct, L1 minor errors, L2 major errors, and L3 complete failure.Each model pair is shown to ten independent annotators.

D.2. Detailed Scoring Logic for Selected Tasks

VBVR-Pro’s selected tasks use verifiable scorers that evaluate final states, intermediate processes, constraints, and scene preservation through task-relevant visual entities. The scoring logic varies by task, using gates and weighted components to penalize invalid trajectories, missing events, and inconsistent outcomes.

  • Scoring framework: Task-specific scorers jointly assess final-state correctness, intermediate-process validity, and preservation of task-irrelevant scene content.The scorers are designed around each task’s required visual entities and temporal behavior.
  • Ball-Cluster Merging: Ball-Cluster Merging combines final-state quality with a process gate, so poor merge ordering, arrivals, counts, or survivor stability suppress the overall score.The final state checks survivor count, compactness, cleanliness, and label accuracy, while the process score evaluates each merge event.
  • Key-Door Navigation: Key-Door Navigation scores trajectory proximity, continuity, and coverage while gating performance by wall-crossing and key-visitation constraints.Skipping the key caps its multiplier at 0.55, and wall hits reduce the corresponding gate.
  • Rolling Ball: Rolling Ball multiplies target-position accuracy by path-following, direction, and intermediate-motion terms, then applies a scene-consistency gate.This prevents a ball that merely ends near the target from receiving a high completion score when its trajectory is invalid.
  • High-Density Liquid and Move Objects to Targets: High-Density Liquid evaluates floating or sinking outcomes alongside physically plausible descent, while Move Objects to Targets rewards transport quality and synchronized motion.Sequential movement drives the synchronization term toward zero and limits the score to 0.8 · movement.
Loading 2608.26105v1…