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

OmniVerifier-C1. Decomposed Graded Verifier for Rule-Governed Generative World Modeling. By Bowei Liu and twelve others, from Tsinghua University, Alibaba Group, and two other institutions. Published on the fifteenth of October. Its code repository is worked out: about one hundred and fifty source files, with a readme.

Generative visual verifiers play a crucial role in advancing multimodal intelligence, and the conventional verifier is binary: it returns True or False and nothing more. The authors set out to make verifiers more robust in complex multimodal scenarios with compositional requirements, where a prompt or a set of environmental rules asks for many things at once and a single discrete verdict is a thin signal to learn from.

OmniVerifier-C1 moves past the binary format. It automatically decomposes a complex prompt, or a set of environmental rules, into fine-grained components, and assigns each component a pointwise score. What comes out is a continuous critique signal that carries more information than a yes or a no. The authors also find that these structured, interpretable outputs improve how well the verifier generalises to out-of-domain multimodal critic tasks. The change is from a verdict to a breakdown. A binary verifier says whether an image passed. A graded, decomposed one says which parts of the request were met, and by how much, and that finer account is what the method is built around.

Training uses a Decoupled RL strategy that optimises two things jointly: binary judgment at the level of each sub-question, and alignment scoring at the global level. The precise local feedback develops fine-grained verification ability, and that ability in turn is the foundation for reliable continuous scoring of the whole. The two levels match the two halves of the verifier's output. Each sub-question gets a yes or no judgment, trained against precise feedback, and the global alignment score is optimised jointly with them, resting on the skill those local judgments build.

Under a general reinforcement learning framework, the pointwise graded verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. The investigation then moves to rule-governed generative world modelling, where generative models are used for world simulation and world reasoning under explicit rules.

Here a second result appears. From a theoretical standpoint, the authors show that when ground-truth images are available, a pairwise verification paradigm yields more accurate critiques than the pointwise formulation. Empirical results on diverse structured world-modelling tasks, such as Maze and Sudoku, further support this. The condition matters. When a correct image exists to compare against, judging a candidate side by side with it beats scoring the candidate alone, and structured worlds like a maze or a sudoku grid are settings where such a reference can exist.

The paper ends on a claim about direction. The move from generative models to structured world-modelling agents, the authors suggest, hinges critically on having accurate pairwise graded visual verifiers. Read in order, the work traces an arc: from binary verdicts, to graded pointwise scores built from decomposed requirements, to graded pairwise comparison wherever a ground-truth image exists to compare against.