OmniVerifier-C1
A generative visual verifier model that proposes a decomposed, hierarchical verification framework. It automatically breaks complex prompts or environment rules into fine-grained sub-questions and produces continuous point-by-point scores, enabling fine-grained evaluation of multimodal generation results and optimization through reinforcement learning.
Timeline (last 90 days)
OmniVerifier-C1 is a decomposed hierarchical verification framework that can automatically break down complex prompts or environment rules into fine-grained sub-questions and give point-by-point continuous scores.
OmniVerifier-C1 adopts a decoupled reinforcement learning training strategy, splitting the optimization objective into two parts: sub-question-level binary judgment and global alignment scoring.
Experiments show that in traditional text-to-image scenarios, the point-by-point hierarchical verifier significantly outperforms the binary verifier.
The paper extends the verifier to rule-governed generative world modeling, a setting in which generative models are used for world simulation and reasoning and must obey explicit rules.
The researchers validated the effectiveness of the hierarchical verification framework using structured world modeling tasks such as Maze and Sudoku.
All Facts (5)
OmniVerifier-C1 is a decomposed hierarchical verification framework that can automatically break down complex prompts or environment rules into fine-grained sub-questions and give point-by-point continuous scores.
50%UnverifiedOmniVerifier-C1 adopts a decoupled reinforcement learning training strategy, splitting the optimization objective into two parts: sub-question-level binary judgment and global alignment scoring.
50%UnverifiedExperiments show that in traditional text-to-image scenarios, the point-by-point hierarchical verifier significantly outperforms the binary verifier.
50%UnverifiedThe paper extends the verifier to rule-governed generative world modeling, a setting in which generative models are used for world simulation and reasoning and must obey explicit rules.
50%UnverifiedThe researchers validated the effectiveness of the hierarchical verification framework using structured world modeling tasks such as Maze and Sudoku.
50%