Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation

Benchmark Model Rank Results
visual-question-answering-vqa-on-hr-bench-4kGLM-4.6V–Overall: 82.13
visual-question-answering-vqa-on-hr-bench-4kGPT-5.2–Overall: 81.12
visual-question-answering-vqa-on-hr-bench-4kGPT-5.4–Overall: 84.00
visual-question-answering-vqa-on-hr-bench-4kGemini-3.1-Pro–Overall: 89.63
visual-question-answering-vqa-on-hr-bench-4kGemini-3.5-Flash–Overall: 89.12
visual-question-answering-vqa-on-hr-bench-4kKimi-K2.6–Overall: 81.88
visual-question-answering-vqa-on-hr-bench-4kMiMo-VL-7B-RL–Overall: 73.50
visual-question-answering-vqa-on-hr-bench-4kMiniCPM-V-4.5–Overall: 69.63
visual-question-answering-vqa-on-hr-bench-4kQwen3-VL-235B-A22B-Instruct–Overall: 86.13
visual-question-answering-vqa-on-hr-bench-4kQwen3-VL-8B-Instruct–Overall: 79.63
visual-question-answering-vqa-on-hr-bench-4kQwen3.5-397B-A17B–Overall: 89.38
visual-question-answering-vqa-on-hr-bench-4kQwen3.5-4B–Overall: 84.38
visual-question-answering-vqa-on-hr-bench-4kQwen3.5-9B–Overall: 85.75
visual-question-answering-vqa-on-hr-bench-4kVision-OPD-4B–Overall: 84.50
visual-question-answering-vqa-on-hr-bench-4kVision-OPD-9B–Overall: 88.13
visual-question-answering-vqa-on-hr-bench-8kGLM-4.6V–Overall: 78.88
visual-question-answering-vqa-on-hr-bench-8kGPT-5.2–Overall: 78.38
visual-question-answering-vqa-on-hr-bench-8kGPT-5.4–Overall: 77.88
visual-question-answering-vqa-on-hr-bench-8kGemini-3.1-Pro–Overall: 86.88
visual-question-answering-vqa-on-hr-bench-8kGemini-3.5-Flash–Overall: 86.62
visual-question-answering-vqa-on-hr-bench-8kKimi-K2.6–Overall: 78.00
visual-question-answering-vqa-on-hr-bench-8kMiMo-VL-7B-RL–Overall: 69.38
visual-question-answering-vqa-on-hr-bench-8kMiniCPM-V-4.5–Overall: 61.50
visual-question-answering-vqa-on-hr-bench-8kQwen3-VL-235B-A22B-Instruct–Overall: 80.38
visual-question-answering-vqa-on-hr-bench-8kQwen3-VL-8B-Instruct–Overall: 75.25
visual-question-answering-vqa-on-hr-bench-8kQwen3.5-397B-A17B–Overall: 85.50
visual-question-answering-vqa-on-hr-bench-8kQwen3.5-4B–Overall: 80.13
visual-question-answering-vqa-on-hr-bench-8kQwen3.5-9B–Overall: 80.63
visual-question-answering-vqa-on-hr-bench-8kVision-OPD-4B–Overall: 80.38
visual-question-answering-vqa-on-hr-bench-8kVision-OPD-9B–Overall: 85.50