ByteDance Releases EdgeBench to Evaluate Long-Horizon Agent Evolution

ByteDance's Seed team released EdgeBench, a framework designed to evaluate AI agents' ability to autonomously iterate and learn in real-world environments over extended periods. Analyzing 38,000 hours of interaction data across 134 tasks, they discovered a log-sigmoid scaling law for performance and exponential improvements in learning speed.

2026-07-06 ~ 2026-07-07 · 4 related posts

• ByteDance EdgeBench: Evaluating AI Agents' 10-Hour Self-Iteration Capabilities — karminski3 · 2026-07-06
• EdgeBench Pt 2: AI Agent Strategy Optimization Boosts Unit Test Score 6x — karminski3 · 2026-07-06
• ByteDance Releases EdgeBench to Measure Agent Learning in Real-World Environments — 字节跳动Seed · 2026-07-07
• EdgeBench: Scaling Laws for Learning in Real-World Environments — ByteDance-Seed · 2026-07-07