[arXiv 2025] SAGE: Training Smart Any-Horizon Agents for Long Video Reasoning with Reinforcement Learning - View it on GitHub