88Z AI, 7 July — ByteDance's Seed team has released EdgeBench to evaluate the continuous learning capabilities of AI Agents in real-world environments. The benchmark includes 134 real tasks across six major domains, supporting extended continuous operation. Research indicates that Agent learning trajectories follow a high-precision log-sigmoid curve, with leading models doubling their learning speed every three months.
At a glance
- EdgeBench includes 134 real tasks across six domains, supporting over 12 hours of continuous operation
- Research finds Agent learning trajectories follow a log-sigmoid curve, with frontier models doubling learning speed every three months
The update
ByteDance's Seed team has officially launched EdgeBench, an ultra-long-horizon benchmark designed to measure the continuous learning performance of AI Agents in real-world scenarios. The benchmark features 134 carefully designed tasks from reality, broadly covering six core areas including scientific exploration, software engineering, and knowledge work. Each task supports continuous operation for over 12 hours to thoroughly test the Agent's long-term interaction and adaptation capabilities. Research based on this benchmark reveals that the learning trajectories of AI Agents during environmental interactions exhibit high-precision log-sigmoid curve characteristics. Furthermore, data indicates a significant improvement in the learning efficiency of frontier models, with their learning speed approximately doubling every three months, demonstrating a trend of rapid evolution.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 字节跳动SeedSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.