Alibaba's ZeroSearch Teaches AI To Search Without Search Engines, Cuts Training Costs By 88%

Alibaba Group researchers have developed "ZeroSearch," a technique that enables large language models to acquire search capabilities without using external search engines during training. The approach transforms LLMs into retrieval modules through supervised fine-tuning and employs a "curriculum-based rollout strategy" that gradually degrades generated document quality. In tests across seven question-answering datasets, ZeroSearch matched or exceeded the performance [PDF] of models trained with real search engines. A 7B-parameter retrieval module achieved results comparable to Google Search, while a 14B-parameter version outperformed it. The cost savings are substantial: training with 64,000 search queries using Google Search via SerpAPI would cost approximately $586.70, compared to just $70.80 using a 14B-parameter simulation LLM on four A100 GPUs -- an 88% reduction. The technique works with multiple model families including Qwen-2.5 and LLaMA-3.2. Researchers have released their code, datasets, and pre-trained models on GitHub and Hugging Face, potentially lowering barriers to entry for smaller AI companies developing sophisticated assistants. Read more of this story at Slashdot.

May 9, 2025 - 04:27

Alibaba's ZeroSearch Teaches AI To Search Without Search Engines, Cuts Training Costs By 88%

Alibaba Group researchers have developed "ZeroSearch," a technique that enables large language models to acquire search capabilities without using external search engines during training. The approach transforms LLMs into retrieval modules through supervised fine-tuning and employs a "curriculum-based rollout strategy" that gradually degrades generated document quality. In tests across seven question-answering datasets, ZeroSearch matched or exceeded the performance [PDF] of models trained with real search engines. A 7B-parameter retrieval module achieved results comparable to Google Search, while a 14B-parameter version outperformed it. The cost savings are substantial: training with 64,000 search queries using Google Search via SerpAPI would cost approximately $586.70, compared to just $70.80 using a 14B-parameter simulation LLM on four A100 GPUs -- an 88% reduction. The technique works with multiple model families including Qwen-2.5 and LLaMA-3.2. Researchers have released their code, datasets, and pre-trained models on GitHub and Hugging Face, potentially lowering barriers to entry for smaller AI companies developing sophisticated assistants.