OmniVerifier-C1 adds graded critique to generative world models

A visual verifier that decomposes a complex prompt into fine-grained components and scores each one on a continuous scale outperforms single True/False judges on text-to-image generation and on rule-governed puzzles like Maze and Sudoku, argues a NeurIPS 2026 paper, OmniVerifier-C1. The authors say moving beyond binary verification gives training a "continuous and more informative critique signal," and that the structured, interpretable outputs improve the verifier's generalization to out-of-domain multimodal critic tasks.

To train the verifier the paper introduces a "Decoupled RL training strategy that jointly optimizes sub-question-level binary judgment and global alignment scoring," using local per-sub-question feedback as the foundation for a reliable global score. The authors also argue, both theoretically and empirically, that a pairwise setup, which compares two candidate images against a ground-truth reference, "yields more accurate critiques than the pointwise formulation when ground-truth images are available."

The framing in the abstract is deliberately strong: "the transition from generative models to structured world-modeling agents critically hinges on the availability of accurate pairwise graded visual verifiers." The abstract itself reports no per-task accuracy numbers; those are left to the paper's tables.