EdgeBench: revealing learning scaling laws from real-world environments
Original title:EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments
AI summary
ByteDance's Seed team released EdgeBench, which contains 134 real-world tasks spanning six domains including scientific discovery and software engineering, with each task supporting at least 12 hours of continuous agent operation. Based on about 38,000 hours of interaction data, it found that overall performance follows a precise log S-shaped scaling law with environment interaction time (R²=0.998), and that agent learning speed doubles about every three months. 51 tasks and the full evaluation framework have been released.
Key Takeaways
- ByteDance's Seed team released EdgeBench,which contains 134 real-world tasks spanning six domains…
- Based on about 38,000 hours of interaction data,it found that overall performance follows a…
- 998),and that agent learning speed doubles about every three months.
- 51 tasks and the full evaluation framework have been released.
Quick Facts
- Published
- Updated
- Category
- Papers & Research
- AI score
- 74
这篇论文首次量化了智能体在134个长周期真实任务中从环境学习的性能曲线,发现log-sigmoid缩放定律精确度极高,且学习速度约每三个月翻倍。对于正在设计长期自主agent的团队,这是一份必须看的底层规律报告。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: EdgeBench: revealing learning scaling laws from real-world environments. July 6, 2026. https://new.nat6.net/en/news/cmra57ei80146ihx8c551d8w6/