Multimodal Meta-Verifier with Explicit Structured Recalibration

Visual outcomes are increasingly central to multimodal large language models, making reliable and fine-grained verification essential for scaling generalist foundation models. In this work, we investigate multimodal meta-verification, which leverages verifier-generated rationales rather than decision-only signals, and explore how to effectively incorporate meta-verification feedback into multimodal verifier training. We identify two key findings. First, symbolic verifier outputs (e.g., bounding boxes) outperform textual explanations as meta-verification rationales, enabling efficient rule-based reinforcement learning rewards while avoiding reliance on model-based rewards from auxiliary judge models. Second, decoupling reinforcement learning objectives for binary judgment and meta-verification substantially outperforms joint reward optimization, due to intrinsic differences in output structure and learning dynamics. Based on these insights, we train OmniVerifier-M1, a generalist visual verifier leveraging symbolic meta-verification and decoupled reinforcement learning. OmniVerifier-M1 provides robust verification and fine-grained error localization, and further enables M1-TTS, a verifier-driven agentic generation system achieving dynamic region-level self-correction. This approach paves the way for more reliable, interpretable, and fine-grained multimodal verification, supporting safer and more controllable foundation model deployment.

Modern AI systems can generate and understand images, but they still struggle to reliably determine whether an image truly matches a user’s request. Current systems often only provide a simple “correct” or “incorrect” answer, without explaining what is wrong or how the image could be improved. In this work, we develop a new visual verification system that can both judge images and identify the exact regions that contain mistakes. Instead of relying on long written explanations, our method uses simple spatial signals, such as pointing to problematic regions in an image. This makes the system faster, more reliable, and easier to train. We also introduce a new training strategy that separately teaches the system how to make accurate decisions and how to explain image errors. This leads to more stable learning and better overall performance. Based on these ideas, we build OmniVerifier-M1, a general-purpose image verifier, and M1-TTS, a system that can iteratively correct images by locating and fixing problematic regions. Our work helps make multimodal AI systems more trustworthy, interpretable, and controllable for future real-world applications.