ByteDance Seed has launched EdgeBench, a new benchmark designed to evaluate AI agents based on their ability to learn through experience while performing real-world tasks over extended periods. This shift in AI evaluation emphasizes not just what agents already know, but how they can adapt and improve in complex environments. EdgeBench includes 134 tasks that agents can work on for 12 to 72 hours, providing informative feedback for continuous improvement. An initial set of 51 tasks, along with the full evaluation framework, has been released to the public to help advance research on long-horizon agent capabilities.

tikgiau: tikgiau is a researcher affiliated with ByteDance Seed who publicly introduced and detailed the EdgeBench benchmark. The account shared key design elements and findings from the project on social media. This directly ties to the news as the source of the announcement and supporting materials.
ByteDance: ByteDance is a global technology company specializing in content platforms and digital services, with active investment in AI research through its Seed division. The company focuses on advancing foundational AI capabilities alongside its consumer-facing products. In this news, ByteDance Seed released EdgeBench to push the boundaries of AI agent evaluation.
EdgeBench: EdgeBench is a benchmark framework for assessing how AI agents improve through extended interaction with real-world tasks and feedback. It emphasizes long-horizon scenarios that simulate practical work environments rather than short, static tests. ByteDance Seed developed and released it to better measure experiential learning in frontier models.

{“Research Release”: “ByteDance Seed has released an initial portion of the tasks and the full evaluation framework to support further development in long-horizon agent research.”, “AI Evaluation Shift”: “The benchmark shifts the focus from agents’ existing knowledge to their capability to learn and improve through continual interaction with complex environments.”}