Today, with the rapid advancement of artificial intelligence technology, how to scientifically measure the continuous learning ability of agents (Agent) in the real world has become a focal point for both academia and industry. Recently, the ByteDance Seed team officially released a long-term benchmark called "EdgeBench," providing a new quantitative reference for research in this field.

The core value of EdgeBench lies in its deep coverage of "learning in real-world environments." This benchmark includes 134 real-world tasks across six fields, and each task requires the agent to work continuously for at least 12 hours. This design aims to break through the limitations of short-term task evaluation, more realistically simulating the long-term performance of agents in complex and dynamic environments. To build this rigorous testing system, the development team collected approximately 38,000 hours of interaction data.

image.png

Research results reveal an interesting trend: the performance of agents in environmental learning follows a high-precision log-sigmoid curve, with an R² of up to 0.998. This means that the learning process of agents shows a strong regularity. Additionally, the data analysis shows that during the period from September 2025 to May 2026, the learning speed of cutting-edge models showed a strong growth trend, doubling every three months.

Currently, EdgeBench has open-sourced 51 of its tasks and the complete evaluation framework for the developer community. Although this benchmark is still in the stage of academic exploration, it is the first to quantitatively describe the long-term environmental learning patterns. For AI researchers, this pattern not only provides a hard metric to evaluate model capabilities but also points the way for future improvements in agents' environmental adaptability and learning efficiency.