Process Rewards with Learned Reliability

Benchmark Model Rank Results
mathematical-reasoning-on-mathverseACA with BetaPRM (Qwen2.5-VL-7B)#22Accuracy: 46.4
multimodal-reasoning-on-mathvisionACA with BetaPRM (Qwen2.5-VL-7B)#53Accuracy: 26.65
multimodal-reasoning-on-mathvistaACA with BetaPRM (Qwen2.5-VL-7B)#85Accuracy: 64.0