rule-verifiable

TRON introduces a scalable online environment for visual reasoning reinforcement learning that generates unlimited diverse training instances with verifiable answers, showing consistent performance improvements across multiple multimodal benchmarks.