The method reduces the cost of training neural networks by almost 90%, writes the South China Morning Post.

The Chinese company launched the open platform ZeroSearch. In it, neural networks are trained in a simulation without interacting with search engines.
A third-party model first generates relevant pages, and then increasingly longer documents, simulating complex search scenarios. This way, the neural network learns to reason and extract the necessary information.

Sending 64,000 queries to Google costs about $586.7, while working in ZeroSearch with a training AI model of 14 billion parameters costs $70.8, meaning costs are reduced by 88%, writes the South China Morning Post. According to the publication, this approach makes the development of search AI models accessible to small teams.
Neural networks trained in ZeroSearch provide relevant answers more often than those that search for information in search engines, the company claims. However, Alibaba does not specify how the models perform in real search conditions after training.

Graph of LLM rewards for relevant answers using ZeroSearch versus the classic method of connecting to a search engine. Source: Alibaba

Developers can use ZeroSearch to train their models by downloading data from GitHub and HuggingFace.