OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
OmniVerifier-C1 grades decomposed visual rules continuously and beats binary verifiers on image and world-model tasks.
OmniVerifier-C1 is a visual verifier that decomposes complex prompts or environmental rules into fine-grained components and assigns continuous pointwise scores instead of binary true/false judgments. It is trained with decoupled reinforcement learning that jointly optimizes local binary judgments and global alignment. The graded verifier outperformed binary verifiers on text-to-image tasks and generalized better to out-of-domain multimodal critic tasks. On rule-governed world modeling such as Maze and Sudoku, pairwise verification was more accurate than pointwise scoring when ground-truth images were available.
- OmniVerifier-C1 decomposes prompts or rules and assigns continuous pointwise scores.
- Decoupled RL jointly trains sub-question binary judgments and global alignment.
- Graded verification beat binary verifiers on text-to-image critic tasks.
- With ground truth, pairwise grading was more accurate on Maze and Sudoku world tasks.
- AI models
- OmniVerifier-C1
From the original article
1 min read at arxiv.orgGenerative visual verifiers play a crucial role in advancing multimodal intelligence.
By Bowei Liu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.