EdgeBench: Measuring Real-World Environment Learning and Discovering a New Scaling Law

ByteDance released EdgeBench, an ultra-long-horizon benchmark for measuring how agents learn from real-world environments, comprising 134 tasks across six domains.

The research finds that agent performance in environment learning follows a log-sigmoid curve with high precision, and that learning speed doubles roughly every three months. 51 tasks and the evaluation framework have been open-sourced.