Bytedance paper says three training steps are sufficient for AI agents
The authors say that with many training phases, no person can tell which step helps. They give an open model as proof.
Bytedance, the company that makes TikTok, and some US universities wrote a paper on how to train AI agents. Agents are AI models that read the web, use code and do tasks with many steps with no help from a person. The paper says that three training steps are sufficient, but the text gives no test numbers.
How it works
Many developers put training phases one after the other. The paper says that it is then not possible to see which step helps. It also says that to do the work again uses more and more money.
The numbers
The text gives no score, no name of the model and no price. The report says that its open model is on the same level as other models. This is a claim, and there is no check of it.
What is not known
The text does not say which three steps the team uses. It does not say the test results of the model. It does not say if a different team did the work again.
Insight: Teams cannot yet measure single steps
In October 2026, Microsoft and City University of Hong Kong made Taste-Bench, a step test. The paper says no person can tell which step helps. The top of 14 models got 59.7 percent, and a random selection gets 25 percent. The 2 facts show that there is no clear test of one step. This reads as a sign that the group that makes the test that all teams use will have control.
Sources
Posted