@allen_ai
🎥 Introducing SAGE, an agentic system for long video reasoning on entertainment videos—sports, vlogs, & more. It learns when to skim, zoom in, & answer questions directly. On our SAGE-Bench eval, SAGE with a Molmo 2 (8B)-based orchestrator lifts accuracy from 61.8% → 66.1%. 🧵 https://t.co/JkGlw3Ad1Z