Alibaba’s ‘ZeroSearch’ lets AI learn to google itself — slashing training costs by 88 percent
1 min read
Summary
Alibaba’s “ZeroSearch” technique incentivises large language models (LLMs) to develop search capabilities without connecting to real search engines.
During training, the system uses a curriculum-based rollout strategy that progressively lowers the quality of generated responses, simulating the effects of a real search engine over time.
ZeroSearch has been shown in experiments to outperform real search engines in terms of both cost and performance, giving developers more control over the training process and reducing dependence on external tech platforms.
The paper claims that training using real search engines, such as Google, via API would cost around 586.70forapproximately64,000searchqueries,comparedtoonly70.80 for a 14B-parameter LLM, highlighting an 88% cost reduction.
ZeroSearch makes the development of advanced AI training more accessible and potentially levels the playing field for smaller AI companies and startups with lesser budgets.