EdgeBench paper reveals log-sigmoid scaling laws for agent learning

rohanpaul_ai · x · 2026-08-29

The paper 'EdgeBench' analyzes 38,000 hours of agent interaction, finding that overall performance in environment learning follows a precise log-sigmoid scaling law (R² = 0.998). Learning speed roughly doubles every three months. EdgeBench covers 134 long-horizon tasks across domains like scientific discovery, software engineering, and math.

Original post →

More from Research

Research channel →