Search
Alibaba has unveiled and open-sourced a formidable model named ZeroSearch, which leverages a reinforcement learning framework to bolster the search capabilities of large language models (LLMs) independently of external search engines. According to experimental data, ZeroSearch outperforms models reliant on actual search engines across various datasets, while significantly cutting down costs. Notably, a ZeroSearch model equipped with 7 billion parameters demonstrates a performance on par with Google Search in question-answering tasks, surpassing it in certain instances. This achievement comes with a remarkable 87.93% reduction in costs. The research team has generously made the relevant code, datasets, and pre-trained models openly available.
-
C114 Communication Network -
Communication Home
