Jihai Zhang

We introduce ComPABench to evaluate VLM compositional reasoning, showing that existing post-training methods struggle, while enhancing vision-text alignment and using progress rewards improves RL-based compositional ability.