Research map / Vision-language and multimodal / Multimodal LLMs

Generative Universal Verifier as Multimodal Meta-Reasoner

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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.

Introduces ViVerBench (3,594 expert-curated true/false visual verification questions across 16 tasks), trains OmniVerifier-7B from Qwen2.5-VL-7B with RL, and proposes OmniVerifier-TTS, a sequential verify-and-edit test-time scaling loop for image generation.

Problem

MLLMs struggle to verify visual outcomes (generated images, tool-returned intermediate images) reliably, falling well short of human performance on visual verification; there is no benchmark or universal generative verifier for this.

Approach

Build a verification benchmark spanning 16 tasks, train a generative verifier with rule-based RL on automatically constructed true/false pairs (prompt-modification and inpainting pipelines), and use the verifier's explanations to drive sequential image refinement in unified multimodal models.

Key results

• OmniVerifier-7B reaches 0.653 overall rule-based accuracy on ViVerBench, up 8.3 points over Qwen2.5-VL-7B (0.570) and above GPT-4o (0.645).
• OmniVerifier-TTS raises Qwen-Image by 3.7 points on T2I-ReasonBench (55.5 to 59.2) and GPT-Image-1 by 2.5 points (76.8 to 79.3).
• Sequential TTS outperforms parallel Best-of-10 TTS on all benchmarks (e.g., GenEval++ overall 0.718 vs 0.693 for Qwen-Image) while using about 47% of the generation time.

Surprising findings

• Human-level visual verification is far from current VLMs; the best models (Gemini 2.5 Pro, Seed 1.5-VL) score about 0.745 rule-based against human 0.932.
• Maze verification data shows minimal cross-task transfer, attributed to the distribution gap between synthetic maze images and natural images.

Limitations

• Generated verification data relies on proprietary models (GPT-5, Seedream 3.0, Seed 1.5-VL) for construction and filtering.
• Maze and robotics verification show limited generalization from other tasks.
• Model-based evaluation depends on a GPT-4.1 judge.

Experimental setup

Datasets: ShareGPT-4o-Image LVIS ViVerBench (open-source augmentation sources) T2I-ReasonBench GenEval++

Benchmarks: ViVerBench T2I-ReasonBench GenEval++

Metrics: rule-based accuracy model-based accuracy (explanation judged) T2I-ReasonBench overall score GenEval++ overall score

Baselines: GPT-4o Gemini 2.5 Pro GPT-5 OpenAI o3 OpenAI o1 OpenAI o4-mini Seed 1.5-VL Qwen2.5-VL-72B InternVL3.5 A28B QwenVL-TTS Best-of-N SD-3-Medium

Builds on: DAPO Qwen2.5-VL SAM 2.1 FLUX.1-dev

Closest papers by meaning

• DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models (ICML 2026 · similarity 0.90)
• ROVER: Benchmarking Reciprocal Cross-Modal Reasoning for Omnimodal Generation (ICLR 2026 · similarity 0.90)
• Breaking Dual Bottlenecks: Evolving Unified Multimodal Models into Self-Adaptive Interleaved Visual Reasoners (ICML 2026 · similarity 0.90)
• VisuRiddles: Fine-grained Perception is a Primary Bottleneck for Multimodal Large Language Models in Abstract Visual Reasoning (ICLR 2026 · similarity 0.89)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.89)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.89)