WebGPT: Improving the factual accuracy of language models through web browsing

According to an OpenAI blog post, the company developed WebGPT, a version of GPT-3 fine-tuned to use a text-based web browser for answering open-ended questions. The model was trained to perform actions like clicking links and scrolling, then synthesize information from multiple pages to produce answers. The approach aims to improve factual accuracy by grounding responses in real-time web content rather than relying solely on pre-training data. The blog notes that WebGPT's answers were preferred by human evaluators over those from the base GPT-3 in various tasks. This research highlights a method for reducing hallucinations in large language models by enabling them to access and cite up-to-date information from the web. For AI workflow builders, WebGPT demonstrates a viable strategy to integrate live data sources into language model pipelines, enhancing reliability for knowledge-intensive applications.