Research map / Vision-language and multimodal / Multimodal LLMs
Multimodal Meta-Verifier with Explicit Structured Recalibration
Xinchen Zhang, Bowei Liu, Jiale Liu, Chufan Shi, Yizhen Zhang, Junhong Liu, Youliang Zhang, Zhiheng Li, Yujiu Yang, Ling Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Investigates multimodal meta-verification, where verifiers produce rationales rather than decision-only signals, and studies how to incorporate meta-verification feedback into multimodal verifier training.
Problem
Reliable, fine-grained verification of visual outcomes from multimodal LLMs, which is needed for scaling generalist foundation models.
Approach
Uses verifier-generated rationales (meta-verification) instead of decision-only signals, and proposes explicit structured recalibration to fold meta-verification feedback into verifier training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decouple binary-judgment and meta-verification rewards instead of joint reward (ablation) | Qwen3-VL-8B verifier · ViVerBench Overall | accuracy: 0.671 -> 0.680 | Improved |
| decouple binary-judgment and meta-verification rewards instead of joint reward (ablation) | OmniVerifier-7B · ViVerBench Overall | accuracy: 0.661 -> 0.668 | Improved |
| decouple rewards instead of joint reward (ablation) | OmniVerifier-7B · ViVerBench Counting | accuracy: 0.640 -> 0.710 | Improved |
| decouple rewards instead of joint reward (ablation) | Qwen3-VL-8B · RefCOCO Overall | accuracy: 0.847 -> 0.866 | Improved |
| decouple rewards instead of joint reward (ablation) | OmniVerifier-7B · ViVerBench Concept Existence Abs.P. (small subtask) | accuracy: 0.513 -> 0.506 | Hurt |
| decouple rewards instead of joint reward (ablation) | Qwen3-VL-8B · ViVerBench Maze | accuracy: 0.458 -> 0.380 | Hurt |
| use rule-based IoU bounding-box meta-reward instead of model-based textual explanation reward (ablation) | Qwen3-VL-8B verifier · ViVerBench Overall | accuracy: 0.670 -> 0.671 | No clear effect |
| use rule-based IoU bounding-box meta-reward instead of model-based textual explanation reward (ablation) | RL verifier training · per-sample reward computation | time (ms): 20.2 -> 0.021 | Improved |
| use bounding-box meta-reward instead of textual explanation (ablation) | Qwen3-VL-8B · per-card GPU memory (GB) | memory: 58.3 -> 49.9 | Improved |
| add OmniVerifier-M1 as verifier in M1-TTS on top of RePlan | RePlan · T2I-CoreBench Overall | score: 0.629 -> 0.690 | Improved |
| add OmniVerifier-M1 verifier to GPT-Image-1.5 | GPT-Image-1.5 · WISE Overall | score: 0.83 -> 0.88 | Improved |
| use OmniVerifier-M1 instead of Qwen3-VL-8B as verifier in M1-TTS with RePlan | RePlan · WISE Time | score: 0.64 -> 0.61 on Time; 0.66 -> 0.71 on Cultural | Mixed |
Closest papers by meaning
- Generative Universal Verifier as Multimodal Meta-Reasoner (ICLR 2026 · similarity 0.89)
- Perception-Aware Policy Optimization for Multimodal Reasoning (ICLR 2026 · similarity 0.88)
- CARE What Fails: Contrastive Anchored-REflection for Verifiable Multimodal Reasoning (CVPR 2026 · similarity 0.88)
- Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.87)
- Meta-TTRL: A Metacognitive Framework for Self-Improving Test-Time Reinforcement Learning for T2I Generation in Unified Multimodal Models (NeurIPS 2026 · similarity 0.87)
- SRPO: Enhancing Multimodal LLM Reasoning via Reflection-Aware Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
