ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning

Benchmark Model Rank Results
mathematical-reasoning-on-mathverseARM-Thinker-7B#25Accuracy: 41.6
multimodal-reasoning-on-mathvisionARM-Thinker-7B#56Accuracy: 25.9
multimodal-reasoning-on-mmmuARM-Thinker-7B#65Accuracy: 57.2
multimodal-reasoning-on-mathvistaARM-Thinker-7B#66Accuracy: 70.2