What happened
Researchers from TikTok owner Bytedance, working with several US universities, have presented Axion 1.0, an AI agent trained in just three phases that performs on par with leading commercial systems from OpenAI and Anthropic in benchmark tests, The Decoder reported. The team has made the model, its training data and its code freely available.
According to the accompanying paper, the authors credit the results to careful selection of training data rather than to piling on additional training steps. They contend that agent training has become heavily layered, which makes it difficult to identify which individual stage actually improves a model and raises the cost for others trying to reproduce published results.
Context
AI agents are increasingly expected to search the web on their own, execute code and complete tasks that involve several steps. The paper describes the methods used to train such capabilities as growing ever more complex, with developers chaining many phases in sequence, a trend it says obscures which techniques matter and makes replicating earlier work expensive.
Le briefing
La semaine tech, en un seul e-mail.
Les dépêches qui méritent votre temps, chaque semaine. Désinscription à tout moment.
