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

The paper presents OmniVerifier‑C1, a graded verifier that is broken into smaller parts to make visual verification more reliable in rule‑based generative world models. Traditional binary verifiers give only simple yes/no answers, which makes it hard for them to handle complex, compositional rules. To solve this, the authors suggest a framework that splits complicated environmental rules or prompts into fine‑grained components. Each component receives a continuous, point‑by‑point score, giving a detailed critique. The system is trained with a Decoupled Reinforcement Learning (RL) strategy that optimizes both the binary decisions for each sub‑question and the overall alignment score at the same time. The paper tests the approach on text‑to‑image generation and

general world‑reasoning tasks. The method could also be useful for civil and environmental engineering. For example, it could verify that generative flood or hydrological models obey physical conservation laws and respect boundary conditions. This resource offers a clear structure and implementation advice for adapting the decomposed verification framework to physical‑world simulations. A key limitation is that the original work focuses mainly on general multimodal and visual tasks. Applying it directly to specialized hydraulic or environmental data will need extra domain‑specific adaptation and validation of the physical rule sets.

Message: Hi FE, I need help with “OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling” in “Civil Engineering”