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

TL;DR

The paper introduces OmniVerifier-C1, a decomposed graded verifier that breaks down complex multimodal prompts into fine-grained components and assigns continuous scores for more nuanced evaluation. It uses a decoupled reinforcement learning approach to train both local binary judgments and global alignment scoring, improving performance in text-to-image and rule-based world modeling tasks. The method outperforms binary verifiers and shows that pairwise verification is more accurate when ground truth is available.

Why It Matters

Accurate, interpretable verification is essential for building reliable generative agents in complex environments. This work advances the development of structured world-modeling systems by providing a scalable, graded critique mechanism that supports better generalization and reasoning under explicit rules.

Key Findings

• A decomposed graded verifier that assigns continuous scores to individual components of complex prompts improves verification accuracy over binary outputs.
• The decoupled RL training strategy enables effective learning of both local sub-question judgments and global alignment, enhancing overall verification robustness.
• Pairwise verification with ground-truth images yields more accurate critiques than pointwise scoring, validating the approach on structured tasks like Maze and Sudoku.