Alibaba has introduced ZeroSearch, an innovative reinforcement learning framework designed to enhance the search capabilities of large models without directly interacting with real search engines. This framework leverages the extensive knowledge these models gather during large-scale pre-training to transform it into a retrieval module that can generate relevant content based on search queries. A unique feature of ZeroSearch is its ability to dynamically control the quality of the generated content, a capability that traditional search engines lack.
In terms of cost efficiency, researchers conducted approximately 64,000 search queries using Google's search via SerpAPI, which cost around $586.70. However, simulating these queries on a large model with 14 billion parameters using four A100 GPUs reduced the cost to just $70.80, achieving a cost reduction of over 80%.
