ZeroSearch: An Innovative LLM Search Engine Framework Open-Sourced by Alibaba 70

On May 8, 2025, Alibaba open-sourced ZeroSearch, a large model search engine. ZeroSearch is an innovative large model search engine framework open-sourced by Alibaba's Tongyi Lab. It is essentially a unique reinforcement learning framework that can fully stimulate the search capabilities latent within large models without requiring interaction with traditional real search engines.

ZeroSearch Main Features

• No Real Search Engine Interaction Required: By simulating the way search engines work, it stimulates the search capabilities of large models, avoiding interaction with real search engines (such as Google) and reducing costs and uncontrollability.

• Dynamic Document Quality Control: Supports the generation of relevant or noisy documents. By adjusting keywords in prompts, it flexibly controls the quality of generated documents, providing diverse retrieval scenarios for training.

• Supports Multiple Models and Algorithms: Compatible with large models of different parameter scales (such as 3B, 7B, 14B), and supports multiple reinforcement learning algorithms (such as PPO, GRPO).

ZeroSearch Core Technologies

• No Real Search Engine Interaction Required: Through its reinforcement learning framework, ZeroSearch can stimulate the search capabilities of large models without interacting with real search engines like Google. It is based on the knowledge accumulated by large models during large-scale pre-training, transforming it into a retrieval module that generates relevant content based on queries.

• Dynamic Document Quality Control: The framework can dynamically control the quality of generated documents, supporting the generation of relevant or noisy documents, providing diverse retrieval scenarios for training. By adjusting keywords in the prompt templates, it guides the model to learn quality-graded retrieval.

• Lightweight Supervised Fine-tuning: First transforms the large model into a retrieval module, enabling it to generate high-quality or low-quality documents to adapt to different training needs.

• Curriculum Learning Mechanism: Gradually increases the document noise level during training, allowing the model to start from simple scenarios and progressively adapt to more challenging tasks, enhancing reasoning capabilities.

• F1 Score-based Reward Mechanism: Uses F1 score as the reward signal, focusing on answer accuracy.

• Multi-turn Interaction Template: Designs clear reasoning, search, and answer stages to guide the model in completing tasks step by step.

Performance and Cost Advantages

• Performance: Researchers tested on seven major question-answering datasets including NQ and TriviaQA. The 7-billion-parameter supervised fine-tuned model achieved a search capability score of 33.06, while the 14-billion-parameter model reached 33.97, surpassing Google's search score of 32.47.

• Cost Efficiency: Training with Google Search for approximately 64,000 search queries costs about $586.70, while using a 14-billion-parameter large model for simulated training on four A100 GPUs costs only $70.80, achieving a cost reduction of over 80%.

ZeroSearch Application Scenarios

• Intelligent Q&A Systems: Can quickly and accurately answer user questions, suitable for scenarios such as intelligent customer service and intelligent assistants.

• Enterprise Internal Knowledge Management: Enterprises can leverage ZeroSearch to build efficient knowledge search platforms, enabling employees to quickly locate the information they need and improve work efficiency.

Project Links

Project Website: https://alibaba-nlp.github.io/ZeroSearch/

GitHub Repository: https://github.com/Alibaba-nlp/ZeroSearch

HuggingFace Model Library: https://huggingface.co/collections/sunhaonlp/zerosearch

arXiv Technical Paper: https://arxiv.org/pdf/2505.04588