Generative Universal Verifier as Multimodal Meta-Reasoner

• Visual outcomes include generated images and intermediate visual states; their high dimensionality and ambiguity make reliable assessment difficult without explicit verification.Source: §1 · p. 2
• ViVerBench evaluates visual-outcome verification across 16 subtasks and six categories, and testing nine state-of-the-art VLMs reveals weaknesses in alignment, world knowledge, and visual reasoning critics.The benchmark contains 3,594 questions manually annotated by 12 domain experts.Source: §1 · p. 2
• OmniVerifier-7B improves ViVerBench by 8.3 points and beats GPT-4o.Source: §1 · p. 2
• Two automated data pipelines support OmniVerifier-7B training, which identifies explicit alignment, relational verification, and integrative reasoning as atomic capabilities that generalize across tasks.Source: §1 · pp. 2–3
• OmniVerifier-TTS iteratively verifies and edits generated images, improving generative quality across reasoning-based and compositional generation and outperforming parallel test-time scaling methods.Validation across two advanced unified multimodal models is reported.Source: §1 · pp. 2–3
• The paper also extends universal verification beyond generation to broader world-modeling scenarios.Source: §1 · p. 3

• Multimodal research has moved from image understanding toward long-chain reasoning and unified models integrating text and image inputs and outputs.Source: §2 · p. 3
• Interleaved reasoning brings visual outcomes into reasoning, while prior work explores tool-assisted reasoning and text-image generation under unified architectures.Source: §2 · p. 3
• The section argues that self-critique can support autonomous learning through verification of textual and visual outcomes without external outcome labels.Source: §2 · p. 3

3.1 Benchmark Overview

• ViVerBench covers 16 tasks across six categories, with true/false labels and explanations for false samples.Source: §3.1 · p. 3

3.2 Benchmark Curation Pipeline

• The four-stage pipeline combines expert-curated, programmatic, and augmented open-source examples with review, human evaluation, and final ambiguity and correctness cleanup.Source: §3.2 · pp. 4–5

3.3 Evaluation Methods

• Rule-based scoring measures answer correctness, while model-based scoring also requires explanation consistency for false cases.Source: §3.3 · p. 5

3.4 Evaluation of Advanced MLLMs

• 0.745 rule-based ViVerBench score for Gemini 2.5 Pro and 0.693 model-based score for Seed 1.5-VL lead, yet remain far from reliable general verification.Source: Table 1 · p. 6, Table 2 · p. 6, §3.4 · p. 6
• About 0.2 separates state-of-the-art VLMs from human performance on three Concept Existence tasks, particularly with overlapping, small, occluded, or blurred attributes.Source: §3.4 · p. 6
• VLMs' extensive world knowledge is not effectively activated in visual verification, with human gaps on both Static and Dynamic Physics tasks.Source: §3.4 · p. 6
• On Maze, FrozenLake, and Robotics, most VLMs perform near chance while humans are near-perfect; Gemini 2.5 Pro scores 0.580 on Maze.Source: §3.4 · p. 7
• Model-based evaluation lowers every model's score versus rule-based evaluation, especially on Dynamic Physics and Maze.Source: Table 1 · p. 6, Table 2 · p. 6, §3.4 · p. 7

4.1 Automated Construction of Visual Verifier Data

• The data construction challenge is to produce difficult, accurate verification examples without overcritical judgments of subjective prompt details.The text identifies attributes and spatial relationships in complex scenes as frequent alignment problems, while atmosphere, style, and fine-grained details may lack uniquely correct answers.Source: §4.1 · p. 7
• The two data pipelines create paired true and false examples by modifying prompts for fixed images or inpainting images for fixed prompts.For inpainting, mask area guides difficulty, and constrained prompts highlight the selected object's bounding box.Source: §4.1 · pp. 7–8

4.2 Generalization of Atomic Capabilities in the Visual Verifier

• Object and spatial verification training generalizes to alignment tasks, while maze training produces no meaningful gains on broader tasks.Object data also transfers to several relational tasks; spatial data gives larger improvements on Non-Spatial, Bounding Box, and Counting, but benefits for Maze and Robotics remain limited.Source: §4.2 · pp. 8–9
• The three atomic capabilities are Explicit Alignment, Relational Verification, and Integrative Reasoning, arranged from perceptual matching through semantic relations to task-level reasoning.They respectively cover directly matchable text-image elements, relationship checks or light reasoning, and holistic prompt-image reasoning.Source: §4.2 · p. 10
• Reinforcement learning supports broad transfer within and between Explicit Alignment and Relational Verification, but Integrative Reasoning shows little cross-task transfer and needs task-specific datasets.A single dataset capturing shared visual patterns is reported as sufficient for broad transfer between the first two capabilities.Source: §4.2 · pp. 9–10

4.3 Omni-capable Generative Universal Verifier

• 8.3% overall performance enhancement in rule based evaluation; OmniVerifier-7B surpasses GPT-4o and reaches capabilities comparable to Qwen2.5-VL-72B.The model was trained on 28k high-quality visual verification examples.Source: §4.3 · p. 10

5.1 Architecture

• The verifier judges a generated image, explains false judgments, and supplies an edit prompt that guides the UMM's fine-grained image edits.Refinement repeats until the verifier returns true or the maximum number of steps is reached.Source: §5.1 · p. 11, Figure 5 · p. 11

5.2 Experiment

• OmniVerifier-TTS improves T2I-ReasonBench by 3.7 points for Qwen-Image and 2.5 points for GPT-Image-1.On compositional generation, the text reports advantages in Color and Pos/Count and says OmniVerifier-TTS leads QwenVL-TTS on both benchmarks.Source: §5.2 · pp. 11–12
• The paper reports iterative refinement correcting unrealistic or flawed image elements, particularly those involving physics or authenticity.The examples are presented as evidence of interleaved image-text generation and reasoning.Source: §5.2 · p. 12, Figure 6 · p. 12

5.3 Comparision between Parallel and Sequential Test-Time Scaling

• Sequential TTS consistently outperforms Parallel TTS across all three benchmarks and achieves superior results in approximately 47% of the generation time.Parallel TTS generates 10 images per prompt; sequential refinement uses verifier critiques for multi-round optimization.Source: §5.3 · p. 12, Table 4 · p. 13

5.4 Extend to Broader World-Modeling Interleaved Reasoning

• Task-specific verifiers provide critiques that correct maze-navigation and block-placement errors in broader world-modeling reasoning scenarios.In the maze setting, the policy can attempt to walk through walls; robotics placement follows a logical order.Source: §5.4 · p. 13, Figure 8 · p. 13

6 Conclusion

• The paper introduces ViVerBench as a comprehensive benchmark for evaluating MLLM verification of visual outcomes.Source: §6 · p. 13
• It develops two automated data-construction pipelines and training strategies that produce OmniVerifier-7B.Source: §6 · p. 13
• It proposes sequential test-time scaling for image generation and explores universal-verifier applications in broader world-modeling reasoning.Future work will scale up the universal verifier and examine its potential for improving multimodal post-training.Source: §6 · p. 13

Concept Existence

• Concept Existence covers object presence and quantities, correct attribute binding, and abstract visual patterns and compositional relationships.Object checks include small, occluded, or confusable objects; attribute checks include color, quantity, and expression; abstract-pattern tasks test visual logic with plausible distractors.Source: A Details of ViVerBench · pp. 18–19

Object Relationship

• Object Relationship evaluates spatial alignment across viewpoints and complex scenes, plus non-spatial interactions not reducible to positions.Non-spatial cases especially include interactions between one object and several others.Source: A Details of ViVerBench · pp. 20–21

World Dynamics

• World Dynamics tests single-image physical realism, including subtle violations of gravity, buoyancy, or lighting, and physical plausibility across image sequences over time.Dynamic evaluation considers temporal consistency and physical laws across a sequence.Source: A Details of ViVerBench · pp. 21–22

Image Annotation

• Image Annotation checks bounding-box correspondence, point correctness, and whether counted-object points identify every target exactly once.Bounding-box tasks emphasize ambiguous object attributes or complex spatial configurations; counting is more demanding than standard pointing because targets cannot be missed or duplicated.Source: A Details of ViVerBench · pp. 22–23

State Value Evaluation

• State Value Evaluation covers valid maze paths, FrozenLake step and outcome correctness, and whether robot block arrangements follow a valid stacking sequence.FrozenLake assesses whether actions are reasonable and reach the target state; robotics scenarios judge intermediate configurations against placement order.Source: A Details of ViVerBench · pp. 24–25

State Value Evaluation; STEM

• GUI tasks assess whether boxes identify actionable interface elements, while STEM tasks check whether image values and variables agree with code.GUI evaluation spans mobile, PC, and web interfaces.Source: A Details of ViVerBench · pp. 25–26

A Details of ViVerBench

• Chart items check statistical plot–code agreement on values, colors, and legends; an example is false because the blue circle’s z-coordinate is 0.25 in code versus 0.75 in the image.Source: A Details of ViVerBench · pp. 26–27
• LaTeX items assess fine-grained code–image correspondence, including values, variables, and operators; an example is false because the code’s denominator is d-2 but the image shows d-1.Source: A Details of ViVerBench · p. 27

B LongCoT Pattern Analysis of OmniVerifier

• With end-to-end reinforcement learning and no cold start, OmniVerifier learns structured chain-of-thought reasoning using format and answer-correctness rewards.The training yields a high-quality, well-structured reasoning pattern.Source: B LongCoT Pattern Analysis of… · p. 27
• Examples identify a missing red lantern and a mismatched smile while reporting that the other listed objects or attributes are correctly represented.The alley example flags the absent lantern; the character example flags the unsmiling expression.Source: B LongCoT Pattern Analysis of… · pp. 28–29
• Decomposing complex verification into simpler subproblems improves accuracy, while binary rule rewards preserve coherent explanations and edit prompts without direct textual-output supervision.Task-specific supervision over explanations is not required to maintain language modeling quality.Source: B LongCoT Pattern Analysis of… · p. 29

C Limitations and Future Works

• Tasks with a large domain gap, such as maze tasks, may generalize less effectively and require task-specific data.The authors identify improving generalization through training and data-construction strategies as future work.Source: C Limitations and Future Works · p. 30
• OmniVerifier-TTS depends on its backbone, whose generated or edited image distribution can produce style artifacts during multi-step refinement.GPT-Image-1, for example, tends to make images increasingly yellowish; the authors say these artifacts affect style, not verification performance.Source: C Limitations and Future Works · p. 30
• The authors distinguish backbone style consistency from verification performance, identifying style consistency under multi-step refinement as an area for further effort.They characterize the issue as a subtle backbone limitation rather than a limitation of OmniVerifier-TTS itself.Source: C Limitations and Future Works · p. 30