SCOUT Agent Improves Ultra-Long Egocentric Video Reasoning.

Understanding extremely long egocentric videos, which can span hours or days, presents significant challenges for current multimodal AI models due to limited context and difficulty in grounding sparse, temporally distributed evidence. While Chain-of-Tool-Thought (CoTT) systems offer iterative retrieval, they often suffer from error propagation because of their rigid "zoom-in" strategies and lack of recovery mechanisms. This research introduces SCOUT, a Self-Checking Chain-Of-Tool-thought framework designed to address these issues. SCOUT incorporates a recovery-aware agentic system with an adaptive policy that evaluates intermediate tool observations. This allows it to dynamically balance between exploiting known information (zooming in) and exploring new regions, enabling more robust multi-hop reasoning over very long time horizons. Furthermore, training such multi-turn tool-using agents is difficult due to sparse rewards and suboptimal credit assignment over long decision trajectories. To overcome this, SCOUT utilizes UPS-GRPO, an uncertainty-prioritized policy optimization method that focuses exploration on high-uncertainty states, improving sample efficiency. It also introduces a turn-level advantage decomposition, combining outcome rewards with tool-grounded temporal alignment rewards for better credit assignment. Experiments show SCOUT achieves state-of-the-art results on ultra-long egocentric benchmarks.