Self-Supervised Visual On-Policy Distillation

Benchmark Model Rank Results
mathematical-reasoning-on-mathverseS^2VOPD (Qwen3.5-9B, Vision-OPD-6K) (extra data)–Accuracy: 75.63
multimodal-reasoning-on-mathvisionS^2VOPD (Qwen3.5-4B, FineVision-12K) (extra data)–Accuracy: 72.66
multimodal-reasoning-on-mathvistaS^2VOPD (Qwen3.5-4B, Vision-OPD-6K) (extra data)–Accuracy: 81.5
visual-question-answering-on-v-benchS^2VOPD (Qwen3.5-4B, FineVision-12K) (extra data)–Accuracy: 91.48