ZeroSearch - A large-scale open-source search engine framework from Alibaba.

What is ZeroSearch?

ZeroSearch is an innovative large-scale model search engine framework open-sourced by Alibaba's Tongyi Lab. It leverages reinforcement learning to incentivize the search capabilities of large models without interacting with real search engines. The framework utilizes pre-trained knowledge from large models, transforming it into a retrieval module that generates relevant or noisy documents based on queries, dynamically controlling the quality of the generated documents. On multiple question-answering datasets, ZeroSearch outperforms Google Search, significantly reducing training costs (by over 80%). Based on lightweight supervised fine-tuning and a curriculum learning mechanism, ZeroSearch progressively improves the model's inference capabilities, supports various reinforcement learning algorithms, and possesses strong scalability and versatility.

Main functions of ZeroSearch

• No real search engine interaction requiredThis approach simulates search engines, incentivizes the search capabilities of large models, avoids interaction with real search engines (such as Google), and reduces costs and unpredictability.
• Dynamically control document qualityIt supports the generation of relevant or noisy documents, and flexibly controls the quality of generated documents based on the keywords in the adjustment prompts, providing diverse retrieval scenarios for training.
• Reduce costsCompared to using real search engines for reinforcement learning training, ZeroSearch significantly reduces training costs (by more than 80%), making large-scale training more feasible.
• Supports multiple models and algorithmsIt is compatible with large models of different parameter sizes (such as 3B, 7B, 14B) and supports a variety of reinforcement learning algorithms (such as PPO, GRPO).

ZeroSearch's technical principles

• Simulated search engineBased on the knowledge of the large model itself, it is transformed into a simulated search engine, which generates relevant or noisy documents according to the query, thus replacing the real search engine.
• Lightweight supervision and fine-tuningFine-tuning a large model based on a small amount of labeled data can generate high-quality or low-quality documents to adapt to different training needs.
• Course learning mechanismDuring training, the noise level of the documents is gradually increased, allowing the model to start with simple scenarios and gradually adapt to more challenging tasks, thereby improving its reasoning ability.
• Reward mechanism based on F1 scoreUsing the F1 score as a reward signal, we focus on the accuracy of the answer and ensure that the answer generated by the model matches the real answer as closely as possible.
• Multi-turn interaction templateDesign clear reasoning, search, and answer phases, and guide the model to complete the task step by step based on structured labels (such as <think>, <search>, <answer>).

ZeroSearch's project address

• Project official 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

Application scenarios of ZeroSearch

• Intelligent question answering systemIt can quickly and accurately answer user questions, and is suitable for intelligent customer service and intelligent assistants.
• Content creationIt helps creators obtain information, generate first drafts, or provide inspiration, and is suitable for news, copywriting, and academic writing.
• Education and LearningProvides students with instant answers, supporting online education and intelligent tutoring.
• Enterprise knowledge managementIt helps employees quickly retrieve internal company resources and improve work efficiency.
• Research and DevelopmentIt provides researchers with the latest research findings and accelerates the research process.