OmniVerifier-C1: Decomposed Graded Verifier for Rule-Governed Generative World Modeling
OmniVerifier-C1 scores decomposed visual rules continuously and outperforms binary verifiers on image and world-modeling tasks.
Researchers introduce OmniVerifier-C1, a visual verifier that splits complex prompts and environmental rules into components and assigns continuous pointwise scores instead of binary True/False judgments. A decoupled reinforcement-learning method jointly optimizes sub-question binary decisions and global alignment. Pointwise grading outperformed binary verifiers on text-to-image tasks, while pairwise verification was more accurate when ground-truth images were available, including Maze and Sudoku world-modeling tests.
- Replaces binary True/False visual checks with decomposed pointwise scores.
- Decoupled RL jointly trains local judgments and global alignment.
- Pointwise grading beat binary verifiers on text-to-image tasks.
- Pairwise grading was more accurate when ground-truth images existed.
- Maze and Sudoku tests supported rule-governed world modeling.
- AI models
- OmniVerifier-C1
From the original article
1 min read at huggingface.co
Generative visual verifiers play a crucial role in advancing multimodal intelligence.
By Bowei Liu, Xinchen Zhang, Xuhuan Li · huggingface.co
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.