BABA(US:BABA) AI科技大本营·2025-05-09 09:35Core Insights - Alibaba's Tongyi team has launched ZeroSearch, a generative search engine framework that operates independently without external search interfaces, achieving low-cost and high-performance retrieval capabilities [1][10]. Group 1: ZeroSearch Overview - ZeroSearch allows users to run a 14 billion parameter model on four A100 GPUs for just $70.80, providing search capabilities that can rival or exceed Google [1][16]. - The framework employs a novel reinforcement learning approach to train search capabilities without interacting with real search engines, addressing issues of document quality and high API costs [2][6]. Group 2: Training Methodology - The training process involves lightweight supervised fine-tuning to convert a large model into a retrieval module capable of generating relevant and irrelevant documents based on queries [8]. - A curriculum learning strategy is introduced, gradually lowering document quality to challenge the model's reasoning and retrieval abilities, thus enhancing its search learning path [2][8]. Group 3: Cost Efficiency and Performance - ZeroSearch has demonstrated an 80%-90% reduction in training costs compared to traditional methods, making it a truly low-cost and high-performance solution for AI search training [10][16]. - In various experimental scenarios, ZeroSearch has achieved performance levels that are equal to or better than models trained with real search engines, with a 7 billion parameter model matching Google search quality and a 14 billion parameter version surpassing it [15][16]. Group 4: Open Source and Accessibility - The researchers have made their code, datasets, and pre-trained models publicly available on GitHub and Hugging Face, promoting accessibility for other researchers and companies [16].