<ul><li>Alibaba's 'ZeroSearch' is a new technique that allows large language models (LLMs) to develop advanced search capabilities without using real search engines during training.</li><li>ZeroSearch uses a reinforcement learning framework that incentivizes LLMs to search without interacting with commercial search engines, leading to significant cost savings by reducing the need for expensive API calls.</li><li>In experiments on question-answering datasets, ZeroSearch surpassed models trained with real search engines and achieved comparable or better performance.</li><li>The cost of training an AI system with ZeroSearch using a simulation LLM on GPUs was found to be 88% lower compared to using real search engines, making advanced AI training more accessible and cost-effective.</li></ul>

Alibaba’s ‘ZeroSearch’ lets AI learn to google itself — slashing training costs by 88 percent

Discover more