Learning Active Perception for Pixel-Space Reasoning via Visual-Intent Stratified GRPO
Open paper
Benchmark
Model
Rank
Results
visual-question-answering-on-v-bench
VIS-GRPO
–
Accuracy: 88.5
visual-question-answering-vqa-on-hr-bench-8k
VIS-GRPO
–
Overall: 81.6
Rank counts only results with a code link.