Unified Spatio-Temporal Token Scoring for Efficient Video VLMs

Benchmark Model Rank Results
video-question-answering-on-tomatoMolmo2 (SigLIP2 So400M + Qwen3-4B) baseline#6Accuracy: 36.5
video-question-answering-on-tempcompass-overallMolmo2 (SigLIP2 So400M + Qwen3-4B) baseline#11Accuracy: 69.9
video-question-answering-on-mvbenchMolmo2 (SigLIP2 So400M + Qwen3-4B) with STTS (30%)#19Avg.: 72.7
video-question-answering-on-mlvuMolmo2 (SigLIP2 So400M + Qwen3-4B) with STTS (30%) + Test-Time Scaling (92 frames)#34Accuracy: 70.5