Paper page - OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling

Source: https://huggingface.co/papers/2610.22355 Authors:

,

,

,

,

,

,

,

,

,

,

Abstract

A graded visual verifier decomposes complex multimodal rules into pointwise scores and uses decoupled reinforcement learning to improve generalization, with pairwise verification further enhancing structured world-modeling accuracy.

Generative visual verifiers play a crucial role in advancingmultimodal intelligence. To enhance the robustness of visual verifiers in complex multimodal scenarios involvingcompositional requirements, we move beyond conventional binary verifiers that produce only True/False discrete judgments and introduce a verifier that automatically decomposes complex prompts or environmental rules into fine-grained components and assignspointwise scoresto provide a continuous and more informative critique signal. Our experiments further demonstrate that these structured and interpretable outputs improve the verifier’s generalization to out-of-domain multimodal critic tasks. To train this verifier, we introduce aDecoupled RLtraining strategy that jointly optimizessub-question-level binary judgmentandglobal alignment scoring, using precise local feedback to develop fine-grained verification capabilities that provide a foundation for reliable continuous scoring at the global level. Under a general reinforcement learning framework, we first empirically demonstrate that the proposed pointwise graded verifier significantly outperforms binary verifiers in traditionaltext-to-imagescenarios. We then extend our investigation torule-governed generative world modelingwhere generative models are leveraged for world simulation and world reasoning under explicit rules. From a theoretical standpoint, we further show that apairwise verificationparadigm yields more accurate critiques than the pointwise formulation when ground-truth images are available, with empirical results on diversestructured world-modelingtasks such as Maze and Sudoku further validating the effectiveness of our approach. Overall, our findings suggest a key insight: the transition from generative models tostructured world-modelingagents critically hinges on the availability of accurate pairwise graded visual verifiers.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2610\.22355

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2610.22355 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.22355 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2610.22355 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.