Join our daily and weekly newspapers for exclusive content on the latest updates and industry-composure AI coverage. learn more
Researcher on Alibaba Group A novel approach has developed that dramatically training can reduce the cost and complexity of the AI system to discover information, the need for expensive commercial search engine APIs completely eliminate.
Technology, called “Zerosearch“The large language model (LLM) allows the model to develop advanced search abilities through a simulation approach rather than interacting with real search engines during the training process. This can save important API expenses to innovation companies, while AI systems offer better control over how to get information.
“Strengthening Learning (RL) training requires frequent rollouts, possibly hundreds of thousands of search requests, which obstruct adequate API expenses and severe scalability,” writes the researchers about them. Paper published on Arxiv this week“To resolve these challenges, we introduce zerosearch, a reinforcement of learning outline that encourages LLM’s search abilities without interacting with the actual search engine.”
Alibaba dropped the zero search only if he hugged
Encourage LLM search capacity without searching pic.twitter.com/qfnijno3lh
– AK (@_akhaliq) May 8, 2025
How Zerosearch Trains AI to search without searching engine
the problem is that Zerosearch The solution is important. Companies developing AI assistants who can search for autonomally information, can withstand two major challenges: the unpredictable quality of documents returned by the search engine during training, and for commercial search engines such as Google, hundreds of thousands of API calls for commercial search engines such as Google.
Alibaba’s approach begins with a mild supervised fine-tuning process, which is to convert an LLM capable of generating both relevant and irrelevant documents in response to a query to convert an LLM into a recover module. During the training of reinforcement learning, the system employs that the researchers say “course-based rollout strategy” that gradually degrades the quality of documents generated.
“Our major insight is that LLM has acquired comprehensive world knowledge during large -scale pretering and is able to generate relevant documents giving a discovery query,” the researchers say. “The primary difference between a real search engine and a simulation LLM lies in the text style of returned material.”
Leaving Google behind a fraction of cost
In extensive experiments Seven questions-relation datasetZerosearch not only matches, but often crosses the performance of trained models with real search engines. Remarkable, A 7B-Permander Recovery Module Google achieved comparable performance of search, while A 14 b-parameter module Even improved it.
Cost savings are sufficient. According to the analysis of the researchers, training using around 64,000 search queries Google search through serpent The cost of only $ 70.80-88% costing a reduction of only $ 70.80–88% using 14B-parameter simulation LLM at four A100 GPU.
“This reinforcement shows the feasibility of using a well -trained LLM as an alternative to real search engines in learning setup,” paper note.
What does it mean for the future of AI development
This success is a major change of how the AI system can be trained. Zerosearch Shows that AI can improve on the basis of external devices such as search engines.
The effect may be sufficient for the AI industry. So far, the advanced AI system often requires expensive API calls for services controlled by large tech companies. Zerosearch changes this equation by allowing AI to search instead of using real search engines.
For small AI companies and startups with limited budget, this approach can level the playground. The high cost of API calls has been a major obstacle for admission in developing sophisticated AI assistants. By cutting about 90%in these costs, zerosearch makes advanced AI training more accessible.
Beyond cost savings, this technology gives more control over the training process to developers. When using the actual search engine, the quality of returned documents is unexpected. With simulated discovery, developers can control it properly what information AI sees during training.
Technology works in many model families, including Qwen-2.5 And Llama-3.2And with both Aadhaar and direction-tuned variants. Researchers have provided their code, dataset and pre-informed models Github And Throat faceAllows other researchers and companies to implement the approach.
Like -big language models develop, such as technology Zerosearch Suggest a future where the AI system can develop rapidly sophisticated abilities through self-simulation rather than relying on external services-changing economics of AI development and reducing dependence on large technology platforms.
The irony is clear: In teaching AI to search without the search engine, Alibaba may have created a technique that makes the traditional search engine less essential for AI development. As these systems become more self-sufficient, the technology landscape can look very different in a few years.
