EdgeBench reveals log-sigmoid scaling laws for agent learning in real-world environments

Researchers analyze approximately 38,000 hours of agent interactions across 134 real-world tasks to uncover scaling laws for post-deployment learning. They find that overall performance follows a log-sigmoid scaling law with high precision (R^2 = 0.998), while learning speed roughly doubles every three months.

• EdgeBench is a suite of 134 real-world tasks featuring ultra-long horizons in domains like scientific discovery and software engineering.
• Each task sustains at least 12 hours of continuous agent operation under rich, multilevel feedback.
• The study provides the first evidence of log-sigmoid scaling for environment learning.
• 51 tasks and the full evaluation framework are publicly released to accelerate research.

This work accelerates the study of how agents learn from real-world experience by providing a standardized benchmark and empirical scaling insights.