Perceptual Projection Pruning: Diversity-Aware Video Token Pruning for Multimodal Large Language Models
Open paper
Benchmark
Model
Rank
Results
spatial-reasoning-on-vsi-bench
P3 (Qwen3-VL-8B, 25% retention)
–
Accuracy: 59.5
video-question-answering-on-mlvu
P3 (Qwen3-VL-32B, 25% retention)
–
Accuracy: 71.1
Rank counts only results with a code link.