OmniVerifier-M1: 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.

📋 한 줄 요약

**[Multimodal Verifier / Meta-Verification]** OmniVerifier-M1이 symbolic meta-verification(예: bounding box)이 textual보다 우수함과 binary judgment·meta-verification reward의 decoupling을 발견 — M1-TTS로 dynamic region-level self-correction 가능.

🎯 핵심 기여도

Visual outcome이 multimodal LLM의 중심이 되며 reliable·fine-grained verification이 generalist foundation model 확장에 essential함을 배경 소개.
Multimodal meta-verification — decision-only signal 대신 verifier-generated rationale을 활용 — 을 연구·meta-verification feedback을 multimodal verifier training에 효과적으로 통합하는 방법 탐색.
Key finding 1: symbolic verifier output(예: bounding box)이 textual explanation보다 meta-verification rationale로 우수 — rule-based RL reward를 efficient하게 가능케 하며 auxiliary judge model의 model-based reward 의존 회피.
Key finding 2: binary judgment와 meta-verification의 RL objective decoupling이 joint reward optimization을 substantially 능가 — output structure·learning dynamics의 intrinsic 차이 때문.

💡 핵심 아이디어

Multimodal verification은 symbolic rationale(예: bounding box)과 decoupled RL objective(binary judgment vs meta-verification)로 구현해야 reliable·interpretable·fine-grained하며, 이로부터 verifier-driven self-correction(M1-TTS) 같은 agentic generation system이 자연스럽게 파생된다.

🔬 기술적 접근법

**방법론**: OmniVerifier-M1 — symbolic meta-verification + decoupled RL.
**핵심 기법**: (1) Symbolic verifier output(bounding box)을 meta-verification rationale로 활용, (2) Rule-based RL reward로 model-based judge 의존 회피, (3) Binary judgment·meta-verification RL objective decoupling, (4) Fine-grained error localization 지원, (5) M1-TTS — verifier-driven agentic generation system으로 dynamic region-level self-correction.

📊 주요 결과

Symbolic 출력(bounding box)이 textual explanation 대비 우수한 meta-verification rationale 입증.
Decoupled RL이 joint reward optimization을 substantially 능가.
OmniVerifier-M1이 robust verification·fine-grained error localization 제공.
M1-TTS의 verifier-driven agentic generation이 region-level self-correction 시연.

💭 의의 및 한계

**의의**: Multimodal verifier 학습의 두 원리(symbolic rationale·decoupled objective) 정립, model-based reward 의존 회피로 학습 안정성·효율, agentic self-correction 시스템 가능성 제시. **한계**: 다양 visual task에서 symbolic 출력 형식의 일반화, decoupled RL의 hyperparameter 튜닝 부담, bounding box 외 symbolic 형식의 확장성.

🚀 실용적 활용

Multimodal LLM의 visual outcome 검증.
Foundation model의 safer·controllable 배포.
Agentic generation의 self-correction 메커니즘.