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

May 9, 2025

—

Source URL: https://slashdot.org/story/25/05/09/0113217/alibabas-zerosearch-teaches-ai-to-search-without-search-engines-cuts-training-costs-by-88?utm_source=rss1.0mainlinkanon&utm_medium=feed
Source: Slashdot
Title: Alibaba’s ZeroSearch Teaches AI To Search Without Search Engines, Cuts Training Costs By 88%

Feedly Summary:

AI Summary and Description: Yes

Summary: Alibaba Group’s “ZeroSearch” technique showcases an innovative approach that enables large language models (LLMs) to develop search capabilities without relying on external search engines, demonstrating significant cost savings and competitive performance.

Detailed Description: Alibaba’s introduction of “ZeroSearch” marks a pivotal advancement in the utilization of large language models for retrieval tasks, with notable implications for AI security, efficiency, and accessibility.

– **Technique Overview**:
– ZeroSearch empowers LLMs to serve as independent retrieval modules by employing supervised fine-tuning techniques.
– It employs a unique curriculum-based rollout strategy that gradually degrades the quality of generated documents, providing a systematic approach to improving search capabilities internally.

– **Performance Findings**:
– In tests spanning seven question-answering datasets, ZeroSearch either matched or exceeded the capabilities of LLMs trained with traditional external search engines.
– A 7 billion-parameter (7B) retrieval module demonstrated performance on par with Google Search, while a more robust 14 billion-parameter (14B) version outperformed Google in various metrics.

– **Cost Efficiency**:
– The financial savings associated with training LLMs using ZeroSearch are considerable. Using Google Search via SerpAPI for 64,000 search queries incurs an estimated cost of $586.70 compared to only $70.80 when using the 14B-parameter ZeroSearch model, showcasing an 88% reduction in training costs.

– **Generalization Across Models**:
– This technique is versatile and compatible with various model families including Qwen-2.5 and LLaMA-3.2, broadening its applicability within the AI sector.

– **Open Source Contribution**:
– Alibaba has made the code, datasets, and pre-trained models publicly available on platforms like GitHub and Hugging Face, potentially democratizing access to advanced AI tools for smaller companies and developers.

The significance of ZeroSearch lies not only in improving the efficiency and performance of LLMs but also in the broader implications for AI deployments in real-world applications where cost, performance, and accessibility are critical factors. This development is especially relevant for compliance professionals seeking to understand how emerging technologies can impact data handling and retrieval within the evolving landscape of AI security.

01 1 2 3 4 5 7 a access accessibility Act advanced AI advancement AI AI security AI tool AI tools Alibaba and API app Application applications Arch as based Bi by C capabilities CI CIA co code companies competitive competitive performance compliance compliance professionals cost cost efficiency cost savings Costs critical cross D data Data Handling dataset datasets de demo deployment developer developers development document DoT dual e efficiency emerging Emerging Technologies end ERP External face fact financial fine fine-tuning fine-tuning techniques for g Gen general generalization generated git GitHub Go Google google search grade Group gs H http HTTPS hugging Hugging Face implications in innovative approach inter intern k l land language language model language models large large language model large language models Large Language Models (LLMs) led Li Link llama llm llms lm M made man metrics Mode model models N no non o OCR of on only open open source contribution OPM ory out over parameter performance platform platforms potential Power pre professionals public Q quality queries question Qwen R rate RCE real real-world applications red reduction retrieval retrieval tasks Ro RSA s saving search search capabilities search engine search engines sec sector security side Sig SoC source source contribution SSO Strategy Supervised Fine supervised fine-tuning system T Task tasks tech techniques technologies test the to tool tools Tor TP trained trained models training training costs trie tuning under up US utilization V val version Wi world world applications x zero ZeroSearch