Alibaba’s “Zerosearch” allows AI to learn Google itself – reduces training costs by 88%

Researcher at Alibaba Group We have developed a new approach that allows us to dramatically reduce the cost and complexity of training AI systems and search for information.

Techniques called “Zerosearch“Instead of interacting with real search engines during the training process, large-scale language models (LLMS) will be able to develop advanced search capabilities via simulation approaches. This innovation saves significant API costs for businesses while better controlling the way AI systems learn to retrieve information.

“Reinforcement Learning” [RL] Training requires frequent rollouts and involves potentially hundreds of thousands of search requests. Paper published on Arxiv this week. “To address these challenges, we introduce Zerosearch, a reinforcement learning framework that encourages LLM search capabilities without interacting with real-world search engines.”

Alibaba dropped Zerosearch with her face hugging

Incentivize LLMS search functionality without searching pic.twitter.com/qfnijno3lh

– ak (@_akhaliq) May 8, 2025

Zerosearch How to Train AI to Search Without a Search Engine

That problem Zerosearch Resolution is important. Companies developing AI assistants that can search information autonomously face two major challenges: The unpredictable quality of documents returned by search engines during training and the extremely high cost of making hundreds of thousands of API calls to commercial search engines like Google.

Alibaba’s approach starts with a lightweight, monitored tweaking process to convert LLM into a search module that can generate both related and unrelated documents according to a query. During reinforcement learning training, the system employs what researchers call a “curriculum-based rollout strategy,” gradually decreasing the quality of the generated documents.

“Our key insight is that LLMS can acquire extensive world knowledge during large pre-training and generate relevant documents with search queries in mind,” the researchers explain. “The main difference between a real search engine and a simulation LLM lies in the text style of the returned content.”

Over Google at some cost

In a comprehensive experiment 7 Question Solving Data SetsZerosearch not only matched, but also often outperformed the performance of models trained in real search engines. Surprisingly, a 7B-Parameter Search Module Achieves performance comparable to Google searches, 14B-Parameter Module And it surpassed that.

Cost reductions are significant. Researchers’ analysis shows training using approximately 64,000 search queries Google Search via Serpapi Using the 14B parameter simulation LLM on four A100 GPUs costs around $586.70, but only costs $70.80.

“This demonstrates the possibility of using well-trained LLM as an alternative to actual search engines in reinforcement learning setups,” the paper states.

What does this mean for the future of AI development?

This breakthrough is a major change in the way AI systems train. Zerosearch It shows that AI can improve without relying on external tools such as search engines.

This impact could be significant for the AI ​​industry. Previously, training in advanced AI systems often required expensive API calls to services managed by large tech companies. Zerosearch modifies this equation by allowing AI to simulate searches instead of using real search engines.

For small businesses and startups with limited budgets, this approach could level the arena. The high cost of API calls has been a major barrier to entry in the development of sophisticated AI assistants. By reducing these costs by almost 90%, Zerosearch makes advanced AI training more accessible.

Beyond cost savings, this technique gives developers more control over the training process. When using a real search engine, the quality of the returned documents is unpredictable. Simulated searches allow developers to control exactly what information the AI ​​is seeing during training.

This technique works with multiple model families Qwen-2.5 and llama-3.2and use both the base and instruction tuning variants. Researchers made available codes, datasets and pre-trained models github and Hugging my faceenabling other researchers and companies to implement the approach.

As large-scale language models continue to evolve, they seem like a technique Zerosearch We propose a future in which AI systems can develop more sophisticated functions through self-simulation, rather than relying on external services. It potentially changes the economics of AI development and reduces dependencies on large technology platforms.

The irony is clear. When teaching AI to search without a search engine, Alibaba may have created a technology that does not require traditional search engines to develop AI. As these systems become more self-sufficient, the technology situation can look very different in just a few years.