OmniVerifier-M1: A Novel Multimodal Meta-Verifier Enhancing Verification Through Structured Recalibration

The study presents OmniVerifier-M1, a multimodal meta-verification system designed to improve the verification process in large language models that incorporate visual outcomes. The authors explore the use of verifier-generated rationales over traditional decision signals, revealing that symbolic outputs, such as bounding boxes, are more effective than textual explanations for training purposes. Additionally, the research finds that separating reinforcement learning goals for binary judgment from those for meta-verification leads to better performance than utilizing a joint reward system. OmniVerifier-M1 utilizes these strategies to achieve fine-grained error localization and supports M1-TTS, a system that allows for dynamic self-correction in generated outputs. This work aims to enhance the reliability and interpretability of multimodal verification systems.