- ZeroSearch dramatically reduces the cost of training AI models through simulated searches, eliminating the reliance on external search engines.
- It uses a supervised reinforcement learning system that improves the recall and reasoning capabilities of LLMs.
- It allows companies and developers to train advanced models at low cost, gaining autonomy and control over the process.
Innovation in the field of artificial intelligence has exploded in recent times, especially regarding large language models (LLMs). One of the most significant advances of 2025 has been ZeroSearch, a technology developed by Alibaba that is shaking up the way these models are trained. What exactly is ZeroSearch, and why is it generating so much excitement in the industry? In this article, we delve into this new methodology, how it works, what advantages it offers over traditional methods, and how it could change the development of AI at all levels.
In tech circles, it's the talk of the town: ZeroSearch promises to reduce AI training costs by a staggering 88% . This efficiency leap, far from being a mere marketing ploy, has profound implications for companies large and small, developers, and, of course, for the advancement of general artificial intelligence.
What is ZeroSearch and where does it come from?
ZeroSearch is a new reinforcement learning technique designed to train language models without relying on real external search engines during the training process. This innovation comes from Alibaba's Tongyi Lab, aiming to solve two common problems in training AI models that use web searches: the high cost of using APIs and the unpredictability of the quality of retrieved documents.
Until now, developing advanced assistants, chatbots, or recommendation engines required sending tens of thousands of queries to search engines like Google through paid services, increasing the cost and limiting scalability, especially for companies with tight budgets.
ZeroSearch changes the game by betting on a system in which the LLM itself learns to simulate the operation of a search engine , generating relevant or even noisy (irrelevant) documents in response to queries and thus allowing training without external interaction.
How does ZeroSearch work? A detailed technical explanation
At the heart of ZeroSearch is a reinforcement learning (RL) framework that eliminates the need for actual internet searches during training. Let's examine step by step how this process is structured according to Alibaba's approach and the comprehensive analyses published on the technique.
1. Lightweight supervised tuning to simulate searches
It all begins with supervised fine-tuning (SFT), in which the LLM is trained to behave like an information retrieval module. Through this tuning, it learns to generate response documents to queries, mimicking the textual style and content type that a real search engine would offer. During this initial phase, interaction paths between the model and a search engine are recorded, establishing logs of queries and retrieved documents.
Successful paths, that is, those that lead to the correct answer, are labeled as positive (useful documents), while those that result in errors or incorrect answers are marked as negative (noisy documents). This differentiation will later help the model understand and reproduce the dynamics of a realistic search, including relevant and less useful results.
2. Role of reinforcement learning with curricular simulation
After supervised tuning, the model moves to the reinforcement training phase, where good practices are reinforced and failures are penalized. Here, the simulated LLM itself acts as a search engine, responding to queries generated by the policy model and returning documents that may be useful or problematic.
The difficulty for the model increases progressively, following a curriculum strategy that slowly degrades the quality of the generated documents. This allows the system to learn first in controlled environments and, as it progresses, to encounter increasingly noisy or complex examples . This approach helps the model develop robust search and reasoning capabilities under realistic conditions.
3. Design of rewards and evaluation metrics
To guide learning, ZeroSearch uses a reward function based on the F1 score , which balances accuracy and recall by considering the word match between the prediction and the correct answer. The goal is to maximize the accuracy of the final answers the model can generate, without worrying too much about formatting, since LLMs typically produce well-formatted text naturally.
4. Multi-turn interaction and reasoning templates
During training, interaction templates are used that divide the process into three stages: internal reasoning (delimited between tags like <think>...</think>), conducting the consultation (<search>...</search>) and response generation (<answer>...</answer>). This way, the model can improve both its ability to formulate relevant queries and to provide well-founded answers.
5. Compatibility and scalability
ZeroSearch is compatible with major language models, such as the Qwen-2.5 family, Qwen-2.5 , LLaMA-3.2, and instruction-tuned or base variants. Furthermore, it can be implemented with different reinforcement algorithms (PPO, GRPO, among others), facilitating its adoption in diverse development environments.
Real-world data: How much does ZeroSearch save and how does it perform?
Experiments conducted by Alibaba and documented in specialized publications and repositories show that ZeroSearch achieves comparable, and even superior, performance to that obtained through traditional commercial search engines . The cost savings are particularly significant:
- Performing 64.000 queries using the Google Search API can cost around $586,70 (approx. €540).
- The same query volume, generated and managed with a 14.000 billion parameter LLM using ZeroSearch, reduces the cost to just $70,80 (about € 65).
- This differential implies a 88% savings in training costs, eliminating dependency on external APIs and allowing for greater scalability.
On the other hand, the quality results are impressive: experiments show that a 7B-parameter retrieval module matches the performance of systems based on Google Search, while with 14B parameters the model even outperforms it in question-and-answer tasks, using both single-hop and complex inference datasets.
Key advantages and impact on the artificial intelligence industry
The arrival of ZeroSearch represents a radical shift in how companies and developers can approach training advanced models :
- Drastic reduction of the economic barrier: Facilitates access to advanced AI techniques for SMEs, startups, and independent developers who were previously held back by the cost of commercial APIs.
- Greater control over trainingBy generating simulated documents, teams can define exactly what information the model receives, adjusting the difficulty and quality to suit their needs.
- Boosting technical autonomy: Minimizes dependence on large foreign technology platforms, promoting the local development of customized AI solutions.
- Adaptability and modularityZeroSearch can be deployed on a variety of models and tailored to different workflows and business requirements.
Differences from previous strategies: RAG, real searches and simulations
Before ZeroSearch, the predominant solution for providing up-to-date and accurate information to LLMs was the use of Retrieval-Augmented Generation (RAG), where the model queries external sources using real-world searches. However, this presents some obvious problems:
- High cost: Continued use of APIs can skyrocket budgets.
- variable quality: Retrieved documents can be very inconsistent depending on the searches and the API itself.
- Legal and privacy limitations: Relying on third-party services involves legal and political risks, especially if you are training with sensitive information.
ZeroSearch eliminates the need to continually refer to external sources, allowing the model to learn to search “within itself” as it simulates the experience of interacting with a search engine.
Impact and real-life applications: from Quark to the democratization of AI
Alibaba has already integrated ZeroSearch into its commercial products. Its Quark application, powered by Qwen models, has seen remarkable improvements in reasoning and accurate responses to complex queries thanks to this technique. But perhaps most importantly, ZeroSearch opens the door for smaller companies to design their own advanced models without the need for expensive external infrastructure.
The research community has access to the code repository, datasets, and pre-trained models on both GitHub and Hugging Face, which is fostering global adoption and experimentation.
What will the future of AI training look like thanks to ZeroSearch?
As these techniques mature, we will see a proliferation of intelligent assistants with advanced search capabilities that don't rely on Google, Bing, or similar services. This opens up new opportunities in education, business, and research, while potentially eroding the dominance of major search engines in the artificial intelligence sector.
For Spain and Europe, this represents the possibility of autonomous growth, reduced technological dependence and costs, and greater strategic control over critical information systems.
The rise of ZeroSearch marks the beginning of a new era in which training artificial intelligence models will cease to be a luxury available to only a few, becoming an accessible, scalable, and increasingly sophisticated tool. By teaching AI to search without leaving its own environment, Alibaba has taken a giant leap toward developing self-sufficient, efficient systems adapted to any need . It's no longer just about reducing costs, but about reinventing the rules of the game for the entire artificial intelligence industry.
