Beyond SFT-to-RL: Pre-alignment via Black-Box On-Policy Distillation for Multimodal RL
Open paper
Benchmark
Model
Rank
Results
mathematical-reasoning-on-mathverse
PRISM+GSPO (Qwen3-VL-8B)
#4
Accuracy: 71.5
multimodal-reasoning-on-mmmu
PRISM+GRPO (Qwen3-VL-8B)
#47
Accuracy: 66.6
Rank counts only results with a code link.