New Metrics for Agent Evolution: ByteDance Seed Launches EdgeBench Benchmark

The ByteDance Seed team has introduced EdgeBench, an ultra-long-range evaluation dataset, focusing on measuring agents' ability to learn continuously in real-world environments. The benchmark covers six fields, including 134 real-world tasks, each requiring agents to complete multi-step interactions and dynamically adapt to environmental changes, providing a quantitative reference for research on long-term autonomous learning.