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

Alibaba's innovative ZeroSearch technique enables large language models (LLMs) to learn advanced search capabilities in simulated environments, bypassing costly commercial APIs. By using a lightweight fine-tuning phase with relevant and irrelevant documents followed by curriculum-based rollouts that degrade document quality, the method achieves impressive performance—matching or outperforming Google Search at reduced costs, even with smaller models sometimes surpassing larger ones. This breakthrough marks a significant shift in AI training, offering cost-effective alternatives with greater control over learning outcomes.