BingGPT: A New Language Model from Microsoft

BingGPT is a new language model developed by Microsoft Research that aims to improve the quality and relevance of web search results. BingGPT is based on the GPT-3 architecture, which uses a deep neural network with 175 billion parameters to generate natural language texts from various inputs.

BingGPT differs from GPT-3 in several ways. First, BingGPT is trained on a large and diverse corpus of web documents, including news articles, blogs, social media posts, reviews, and more. This allows BingGPT to capture a wide range of topics, styles, and tones that are relevant for web search users.

Second, BingGPT is fine-tuned on specific tasks related to web search, such as query rewriting, answer generation, summarization, and content creation. This enables BingGPT to produce more accurate and useful texts that match the user’s intent and query.

Third, BingGPT is integrated with Bing’s existing search engine infrastructure, which provides additional information and feedback to enhance the language model’s performance. For example, BingGPT can leverage Bing’s knowledge graph, entity recognition, spell correction, and ranking algorithms to generate more coherent and relevant texts.

BingGPT is currently being tested on a subset of Bing’s web search queries and results. The initial feedback from users and evaluators has been positive, indicating that BingGPT can improve the user experience and satisfaction of web search. Microsoft Research plans to continue developing and improving BingGPT, as well as exploring its applications in other domains and scenarios.