OmniVerifier-M1: A Multimodal Universal Verifier Based on Symbolic Meta-Verification and Decoupled Reinforcement Learning

This paper proposes the OmniVerifier-M1 multimodal verifier, which uses symbolic outputs (e.g., bounding boxes) as the basis for meta-verification and adopts decoupled reinforcement learning objectives. It achieves robust verification capabilities, fine-grained error localization, and dynamic region-level self-correction. This verifier supports general visual verification tasks and can also empower generation systems (e.g., M1-TTS) to improve output quality, providing a foundation for the reliable deployment of multimodal models.

With the development of multimodal large language models, verifying the reliability of generated content has become a key bottleneck for large-scale applications. Traditional verification methods only provide binary judgments, lacking fine-grained localization and interpretability. Meta-verification uses the verifier's own reasoning basis, but how to effectively use meta-verification feedback to train better multimodal verifiers remains an open problem.

Advantages of Symbolic Outputs: Experiments show that symbolic outputs (e.g., bounding boxes) as meta-verification basis are superior to text explanations, due to supporting rule-based rewards, reducing reliance on auxiliary models, and providing precise spatial information.

Decoupled RL Objectives: Decoupling binary judgment and meta-verification objectives is better than joint optimization, avoiding interference from output structure differences, learning dynamic conflicts, and gradient issues.

Based on this, the OmniVerifier-M1 architecture is designed: symbolic meta-verification (output structured symbols) + decoupled reinforcement learning (optimize verification judgment head and meta-verification head separately) + general visual verification capabilities.

Fine-Grained Error Localization: Precisely locate problem areas via bounding boxes, enhancing interpretability and supporting dynamic correction.

M1-TTS System: A verifier-driven generation system that achieves dynamic region-level self-correction, iterative optimization loops, and controllable generation.

Application Prospects: Can be used in scenarios such as image generation verification, visual question answering verification, document understanding verification, content moderation, and medical image analysis.

Experimental Verification: Excellent performance in standard verification task accuracy, meta-verification quality (high scores in human evaluation), error localization precision (superior to text-based methods), and generation quality improvement (significant improvement in M1-TTS).

Technical Contributions: 1. Symbolic meta-verification paradigm; 2. Decoupled RL objective design; 3.U niversal visual verifier OmniVerifier-M1; 4. Demonstration of a verifier-driven generation system.

Limitations: 1. Currently focused on visual verification; expansion to other modalities needs exploration; 2. Symbolic outputs are limited to bounding boxes; more rich structured representations are needed; 3. Decoupled training increases complexity; efficient joint training needs research.

Future Directions: Expand to more modalities and tasks, explore richer symbolic representations, develop efficient end-to-end training methods, and integrate into more generation systems.