Chenyang Gu

Verifiable Rewards Improve Native Visual Reasoning Training

A 300-task procedural suite replaces preference-only judging with task-specific scorers, raising matched reinforcement-learning performance from 0.509 to 0.548.