Generative Universal Verifier as Multimodal Meta-Reasoner

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Generative universal verifier enables VLM meta-reasoning and self-refinement

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Current VLMs lack reliable visual-outcome verification — the ability to judge whether generated or intermediate images satisfy a given prompt or constraint. This paper introduces OmniVerifier, a generative universal verifier trained via reinforcement learning on Qwen2.5-VL-7B, and a sequential test-time scaling paradigm (OmniVerifier-TTS) that iteratively refines generated images through verification-and-editing loops.

Problem

VLMs scale reasoning over text but treat vision as static context; they cannot reliably verify visual outcomes (generated images, intermediate reasoning states in interleaved generation). Without trustworthy visual verification, self-correction, iterative refinement, and test-time scaling for image generation remain infeasible. No comprehensive benchmark or strong open model existed for this capability.

Method

ViVerBench: A manually curated benchmark of 3,594 binary true/false verification questions across 16 subtasks in 6 categories (Concept Existence, Object Relationship, World Dynamics, Image Annotation, State Value Evaluation, STEM). Each question requires a binary judgment plus explanation; two complementary metrics (rule-based and model-based) assess both decision accuracy and explanation validity.

Data construction: Two automated pipelines generate paired true/false training data at scale: 1. Image-Fixed, Prompt-Modified — GPT-5 recaptions complex images with faithful prompts (true), then modifies attributes/relations to produce false variants with explanations. 2. Prompt-Fixed, Image-Inpainting — SAM 2.1 segments objects; FLUX.1-dev inpaints selected regions to create false images; GPT-5 generates prompts anchored to bounding boxes. Difficulty is controlled by mask area. Seed1.5-VL filters samples to Best-of-10 accuracy ≥ 0.6.

OmniVerifier-7B: DAPO (rule-based RL with 9:1 correctness-to-format reward) applied to Qwen2.5-VL-7B for 100 steps on 64 × A100-80G GPUs. Ablation training reveals three atomic capabilities: explicit alignment, relational verification, and integrative reasoning, which generalize and facilitate each other cross-task.

OmniVerifier-TTS: Sequential test-time scaling for unified multimodal models (UMMs). Starting from an initial generated image, OmniVerifier-7B verifies the output and directs targeted editing; the loop repeats iteratively. This contrasts with parallel TTS (Best-of-N), which selects among independent samples.

Key Contributions

• ViVerBench: 3,594-question, 16-task benchmark evaluating visual-outcome verification; reveals three systematic failure modes in current VLMs.
• OmniVerifier-7B: First omni-capable generative verifier; achieves +8.3 points on ViVerBench over Qwen2.5-VL-7B base and surpasses GPT-4o; trained using only atomic-skill data with cross-task generalization.
• Atomic capability analysis: Identifies three irreducible capabilities (explicit alignment, relational verification, integrative reasoning) and shows training on a minimal subset of atomic tasks is sufficient for broad generalization.
• OmniVerifier-TTS: Sequential TTS paradigm for UMMs; outperforms Best-of-N parallel TTS on two generation benchmarks with reduced inference time.

Results

• Gemini 2.5 Pro is the best existing model on ViVerBench (overall 0.745) vs. human performance of 0.932 — gap of ~0.187.
• OmniVerifier-7B scores 0.653 vs. Qwen2.5-VL-7B base at 0.570 (+8.3 pp), beating GPT-4o (0.645 overall).
• OmniVerifier-TTS achieves +3.7 on T2I-ReasonBench and +4.3 on GenEval++ over the base UMMs.
• OmniVerifier-TTS outperforms Best-of-N (parallel TTS) on both benchmarks while reducing inference time.
• Best existing model on Maze: Gemini 2.5 Pro scores 0.580 vs. human 0.997 — near-chance performance on reflective visual reasoning.

Limitations

• OmniVerifier-7B, despite +8.3 overall, shows only modest gains on some categories (Abstract Patterns: 0.514; Non-Spatial: 0.679) and instability on complex visual reasoning tasks (Maze, Robotics) when trained from atomic skills alone.
• Atomic-skill generalization to visual reasoning tasks (Maze, Robotics) is unreliable, suggesting the three atomic capabilities are insufficient to cover reflective/deductive reasoning over rule-based state transitions.
• Model size is fixed at 7B; no scaling analysis to larger verifier sizes is provided in the available text.
• Data construction pipelines rely on GPT-5 and Seedream 3.0, introducing potential proprietary-model dependencies for reproduction.

Relevance to Vision-Language Models

This work directly targets a fundamental gap in VLM self-improvement: the inability to verify their own visual outputs, which blocks iterative refinement during both inference and training. The identification of three atomic verification capabilities provides a minimal training recipe applicable to any VLM backbone, and the sequential TTS paradigm offers a drop-in framework for improving unified generation models. For researchers tracking VLMs, the ViVerBench benchmark establishes a concrete evaluation axis — visual-outcome verification — currently absent from standard VLM leaderboards, and OmniVerifier-TTS demonstrates a practical path to closing the gap between VLM reasoning and controllable generation.