EdgeBench: Scaling Laws for Learning in Real-World Environments

ByteDance-Seed · hf · 2026-07-07

ByteDance's Seed team analyzed approximately 38,000 hours of real-world agent interaction data across 134 diverse tasks. They discovered that performance follows a log-sigmoid scaling law accompanied by an exponential increase in learning speed, revealing the scaling patterns of learning in real-world environments.

Related event: ByteDance Releases EdgeBench to Evaluate Long-Horizon Agent Evolution(4 posts)→

Original post →

More from Research

Research channel →