OmniVerifier-M1: Multimodal Meta-Verifier with Explicit Structured Recalibration

Problem
This paper addresses the gap in reliable verification mechanisms for multimodal large language models (LLMs), particularly in the context of visual outcomes. The authors highlight the need for fine-grained verification to enhance the scalability and safety of generalist foundation models. The work is presented as a preprint and has not undergone peer review.

Method
The core technical contribution is the development of OmniVerifier-M1, a multimodal meta-verifier that utilizes symbolic outputs (e.g., bounding boxes) as meta-verification rationales instead of traditional textual explanations. This approach allows for the implementation of rule-based reinforcement learning (RL) rewards, circumventing the dependency on model-based rewards from auxiliary judge models. The authors propose a decoupled reinforcement learning framework, separating the objectives for binary judgment and meta-verification, which leads to improved performance due to the distinct learning dynamics and output structures involved. The training process leverages a dataset of multimodal inputs, although specific details regarding the dataset size and training compute are not disclosed.

Results
OmniVerifier-M1 demonstrates significant improvements over baseline models in verification tasks. The paper reports that the symbolic verifier outputs yield higher accuracy and robustness in error localization compared to traditional methods. While specific numerical results are not provided in the abstract, the authors claim substantial effect sizes, indicating that the decoupled RL approach outperforms joint reward optimization strategies. The introduction of M1-TTS, a verifier-driven agentic generation system, further enhances the model’s capabilities, allowing for dynamic region-level self-correction.

Limitations
The authors acknowledge that their approach may be limited by the reliance on symbolic outputs, which may not generalize across all multimodal tasks. Additionally, the decoupling of RL objectives, while beneficial, could introduce complexity in training and deployment. The paper does not address potential scalability issues when applying the method to larger datasets or more complex multimodal scenarios.

Why it matters
The implications of this work are significant for the deployment of multimodal foundation models. By providing a more reliable and interpretable verification mechanism, OmniVerifier-M1 enhances the safety and controllability of these models in real-world applications. The findings suggest that leveraging symbolic outputs and decoupled learning can lead to more effective verification strategies, paving the way for future research in multimodal AI systems and their applications in critical domains such as autonomous systems and interactive AI.

Authors: Xinchen Zhang, Bowei Liu, Jiale Liu, Chufan Shi, Yizhen Zhang, Junhong Liu, Youliang Zhang, Zhiheng Li et al.
Source: arXiv:2605.28805
URL: https://arxiv.org/abs/2605.28805v1