Researchers have proposed a new framework called AVA-Encoder for learning video representations that are directly usable by artificial agents. The framework transforms videos into structured knowledge graphs and then reconstructs them back into video. This allows agents to easily understand, query, and manipulate the content of the video. Experiments show that AVA-Encoder outperforms existing methods in several tasks, including policy-only settings where it uses significantly fewer system-prompt tokens. --- Why it matters: This matters because current AI systems lack a way to effectively learn from high-quality human films, limiting their ability to produce cinematic-grade videos. AVA-Encoder addresses this challenge by providing a structured video representation that can be used for agentic reasoning and manipulation. Source: https://arxiv.org/abs/2608.12313