ByteDance launches ultra-long-range benchmark collection EdgeBench

ByteDance's Seed team has launched EdgeBench, an ultra-long-range evaluation dataset used to measure the continuous learning ability of AI agents in real-world environments. The dataset includes 134 real-world tasks covering six major fields, including science, software engineering, and knowledge work, with each task supporting over 12 hours of continuous operation. Research has found that the learning trajectory of agents in environmental interactions follows a high-precision log-sigmoid curve, and the learning speed of cutting-edge models doubles every three months.