ByteDance Seed's EdgeBench benchmark evaluates AI agent learning over 12 to 72-hour tasks, revealing a log-sigmoid scaling law with R² of 0.998 across

ByteDance Seed's EdgeBench benchmark evaluates AI agent learning over 12 to 72-hour tasks, revealing a log-sigmoid scaling law with R² of 0.998 across

Researchers at the Chinese tech giant have been analysing how fast AI agents can improve by performing real-world tasks.