Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generative Universal Verifier as Multimodal Meta-Reasoner

Xinchen Zhang, Xiaoying Zhang, Youbin Wu, Yanbin Cao, Renrui Zhang, Ruihang Chu, Ling Yang, Yujiu Yang

TL;DR

This paper introduces the Generative Universal Verifier to enable reflection and refinement in multimodal reasoning, addressing a critical gap in reliable visual verification. It presents ViVerBench, a challenging 16-task benchmark across 6 categories to systematically assess visual verification in MLLMs, and demonstrates substantial gaps relative to humans. To close this gap, the authors train OmniVerifier-7B via two automated data-construction pipelines, uncovering three atomic capabilities (Explicit Alignment, Relational Verification, Integrative Reasoning) and achieving notable gains on ViVerBench. They further propose OmniVerifier-TTS, a sequential test-time scaling paradigm that iteratively refines image generation and editing within unified models, delivering performance boosts on T2I-ReasonBench and GenEval++ and outperforming parallel scaling methods. Together, these components push toward more trustworthy and controllable multimodal reasoning by integrating universal visual verification into both inference and generation-time processes and extending to broader world-modeling scenarios.

Abstract

We introduce Generative Universal Verifier, a novel concept and plugin designed for next-generation multimodal reasoning in vision-language models and unified multimodal models, providing the fundamental capability of reflection and refinement on visual outcomes during the reasoning and generation process. This work makes three main contributions: (1) We build ViVerBench, a comprehensive benchmark spanning 16 categories of critical tasks for evaluating visual outcomes in multimodal reasoning. Results show that existing VLMs consistently underperform across these tasks, underscoring a substantial gap from human-level capability in reliable visual verification. (2) We design two automated pipelines to construct large-scale visual verification data and train OmniVerifier-7B, the first omni-capable generative verifier trained for universal visual verification and achieves notable gains on ViVerBench(+8.3). Through training, we identify three atomic capabilities in visual verification and demonstrate how they generalize and interact synergistically. (3) We propose OmniVerifier-TTS, a sequential test-time scaling paradigm that leverages the universal verifier to bridge image generation and editing within unified models, enhancing the upper bound of generative ability through iterative fine-grained optimization. Beyond generation, we extend universal verifier to broader world-modeling interleaved reasoning scenarios. Empirically, OmniVerifier-TTS achieves improvements on T2I-ReasonBench(+3.7), and GenEval++(+4.3), outperforming existing parallel test-time scaling methods, such as Best-of-N. By endowing multimodal reasoning with reliable visual verification, OmniVerifier advances both reliable reflection during generation and scalable test-time refinement, marking a step toward more trustworthy and controllable next-generation reasoning systems.

Generative Universal Verifier as Multimodal Meta-Reasoner

TL;DR

This paper introduces the Generative Universal Verifier to enable reflection and refinement in multimodal reasoning, addressing a critical gap in reliable visual verification. It presents ViVerBench, a challenging 16-task benchmark across 6 categories to systematically assess visual verification in MLLMs, and demonstrates substantial gaps relative to humans. To close this gap, the authors train OmniVerifier-7B via two automated data-construction pipelines, uncovering three atomic capabilities (Explicit Alignment, Relational Verification, Integrative Reasoning) and achieving notable gains on ViVerBench. They further propose OmniVerifier-TTS, a sequential test-time scaling paradigm that iteratively refines image generation and editing within unified models, delivering performance boosts on T2I-ReasonBench and GenEval++ and outperforming parallel scaling methods. Together, these components push toward more trustworthy and controllable multimodal reasoning by integrating universal visual verification into both inference and generation-time processes and extending to broader world-modeling scenarios.

Abstract

We introduce Generative Universal Verifier, a novel concept and plugin designed for next-generation multimodal reasoning in vision-language models and unified multimodal models, providing the fundamental capability of reflection and refinement on visual outcomes during the reasoning and generation process. This work makes three main contributions: (1) We build ViVerBench, a comprehensive benchmark spanning 16 categories of critical tasks for evaluating visual outcomes in multimodal reasoning. Results show that existing VLMs consistently underperform across these tasks, underscoring a substantial gap from human-level capability in reliable visual verification. (2) We design two automated pipelines to construct large-scale visual verification data and train OmniVerifier-7B, the first omni-capable generative verifier trained for universal visual verification and achieves notable gains on ViVerBench(+8.3). Through training, we identify three atomic capabilities in visual verification and demonstrate how they generalize and interact synergistically. (3) We propose OmniVerifier-TTS, a sequential test-time scaling paradigm that leverages the universal verifier to bridge image generation and editing within unified models, enhancing the upper bound of generative ability through iterative fine-grained optimization. Beyond generation, we extend universal verifier to broader world-modeling interleaved reasoning scenarios. Empirically, OmniVerifier-TTS achieves improvements on T2I-ReasonBench(+3.7), and GenEval++(+4.3), outperforming existing parallel test-time scaling methods, such as Best-of-N. By endowing multimodal reasoning with reliable visual verification, OmniVerifier advances both reliable reflection during generation and scalable test-time refinement, marking a step toward more trustworthy and controllable next-generation reasoning systems.
Paper Structure (37 sections, 2 equations, 8 figures, 4 tables)

This paper contains 37 sections, 2 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. ViVerBench: Assessment of MLLMs on Visual Verification
  4. Benchmark Overview
  5. Benchmark Curation Pipeline
  6. Initial Dataset Construction
  7. Expert Review and Difficulty Enhancement
  8. Human Evaluation
  9. Dataset Refinement and Final Selection
  10. Evaluation Methods
  11. Rule-based Evaluation
  12. Model-based Evaluation
  13. Evaluation of Advanced MLLMs
  14. Weakness in Fine-Grained and Challenging Image-Prompt Alignment
  15. Mismatched Representation of World Knowledge
  16. ...and 22 more sections

Figures (8)

  • Figure 1: ViVerBench Overview. We show only the false data samples to better highlight data difficulty.
  • Figure 2: Automated pipeline for visual verifier data construction.
  • Figure 3: Training progress of four visual verification tasks on ViVerBench.
  • Figure 4: Atomic capabilities in visual verifier.
  • Figure 5: OmniVerifier-TTS: sequential test-time scaling pipeline of unified multimodal models.
  • ...and 3 more figures