Visual verifiers stop answering only true or false: OmniVerifier-C1 splits prompts into sub-questions and lifts ViVerBench overall to 0.684
Synopsis
Visual verifiers have long returned only a true-or-false verdict, which cannot separate partially correct outputs from worse ones. OmniVerifier-C1 decomposes a complex prompt or environmental rule into fine-grained sub-questions, judges each one binarily, and turns the satisfied fraction into a continuous score, while a decoupled reinforcement-learning scheme optimizes sub-question judgments and global alignment scoring together. On ViVerBench, the Qwen3-VL-8B version trained with graded supervised fine-tuning plus graded reinforcement learning reaches an overall score of 0.684, above 0.671 for binary reinforcement learning alone.
Interpretation
The verifier's output changes from a single true-or-false verdict to a continuous score computed from the fraction of satisfied sub-questions, so partially correct generated samples no longer receive the same evaluation as worse ones. Earlier visual verifiers gave only a binary conclusion on whether an image satisfies a prompt, collapsing imperfect samples of differing quality onto the same negative label. On ViVerBench, a multimodal reasoning and critique benchmark, Qwen3-VL-8B with graded supervised fine-tuning plus graded reinforcement learning reaches an overall score of 0.684, versus 0.671 for binary reinforcement learning alone and 0.672 for binary supervised fine-tuning followed by binary reinforcement learning.
Training turns one sample into two kinds of reinforcement-learning samples, one learning the global alignment score and the other learning a binary answer for a single sub-question, and the two are mixed during optimization. The global score is aggregated from sub-question judgments, so the accuracy of local judgments directly determines whether the continuous score is reliable. On 200 held-out binary sub-question samples, sub-question judgment accuracy after graded reinforcement learning is 94.5 against GPT-5.5 annotations, 96.0 against Gemini-3.1-Pro annotations, and 94.0 against human annotations, versus 76.0, 76.5, and 74.5 after graded supervised fine-tuning.
In rule-governed world-modeling tasks, the verifier takes the ground-truth target image as an extra input and compares directly, instead of first inferring the target state itself. Pointwise verification must derive an implicit target state from the initial state and the rules, so reasoning error and judgment error both enter the score. Across six game tasks, Maze, Sudoku, auxiliary line construction, sequence prediction, word search, and 3D projection, with Qwen-Image-Edit-2509 as the generative backbone, supervised fine-tuning followed by pairwise reinforcement learning scores 85.31 on normal Maze versus 67.81 for pointwise reinforcement learning, and 89.97 versus 77.74 on normal Sudoku.
Perspective
World-modeling evaluation is confined to rule-explicit game environments where the target image is uniquely determined by the initial state and the rules, covering Maze, Sudoku, auxiliary line construction, sequence prediction, word search, and 3D projection plus their hard variants; text-to-image evaluation uses GenEval and CLIPScore, and verifier generalization is evaluated on ViVerBench.
The advantage of pairwise verification depends on a ground-truth image being available, and whether it holds in open-ended settings with non-unique targets has not been tested; more sub-questions are not always better, with mean alignment error 0.039 at 11–20 sub-questions versus 0.072 at 21–30, since excessive decomposition introduces redundant questions and can break the required output format.