News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnWebGPT

WebGPT

A GPT-3 model fine-tuned to answer questions while browsing the web

WebGPT is a fine-tuned version of GPT-3 that answers long-form questions by searching and navigating the web through a text-based browser. OpenAI researchers trained it on the ELI5 dataset using imitation learning and human feedback, publishing the work in December 2021. Its answers were preferred over human demonstrators' and top Reddit answers.

Share on WhatsApp

  • Fine-tunes GPT-3 to answer long-form questions using a text-based web browser
  • Trained by imitation learning, then improved with a reward model from human preferences
  • Best model's answers preferred 56% of the time over human demonstrators
  • Preferred 69% of the time over the highest-voted Reddit answer
  • Trained and evaluated on ELI5, questions from Reddit users

What it is

WebGPT is a question-answering system built by fine-tuning GPT-3, OpenAI's large language model. It works inside a text-based web-browsing environment, so it can run searches and move between pages before answering. The task was set up so that humans could do it too, which made it possible to train the model by copying human demonstrations and then improve it with human feedback. Answers come with references collected during browsing, so readers can check factual accuracy. The team trained and tested it on ELI5, a dataset of questions that Reddit users asked. The work was published as an arXiv paper in December 2021.

How it came about

The project was carried out by OpenAI researchers including Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang and John Schulman, with a total of eighteen authors. The paper was submitted to arXiv on 17 December 2021 and revised twice, with the last revision on 1 June 2022. It appeared under Computation and Language, Artificial Intelligence and Machine Learning.

Inside the model

The model starts from GPT-3 and is fine-tuned on a browsing task written in text, letting it search the web and follow links. Because humans can perform the same task, the researchers used imitation learning, training on demonstrations from human operators. They then optimised answer quality with human feedback: a reward model was trained to predict which answer people preferred, and the best model was produced by behaviour cloning followed by rejection sampling against that reward model. To keep evaluation of factual accuracy tractable, the model had to gather supporting references while it browsed.

What it can do

On ELI5, the strongest WebGPT model gave answers that human evaluators preferred 56% of the time over those written by the human demonstrators, and 69% of the time over the highest-voted Reddit answer for the same question. The requirement to cite sources was designed to make it easier for people to judge whether the facts in an answer were right.

How to use it

The work was published as an arXiv preprint, with versions posted in December 2021, March 2022 and June 2022. The paper lists links to code, data and media through services such as Hugging Face and Replicate.

Safety and testing

Requiring the model to collect references while browsing was intended to make human evaluation of factual accuracy easier.

Reception and debate

Human evaluators preferred WebGPT's answers to those of the human demonstrators 56% of the time, and to the top-voted Reddit answers 69% of the time, on the ELI5 questions tested.

Why it matters

WebGPT combined web browsing with a language model trained on human feedback, an approach that fed into later OpenAI work on assistants that can look things up and cite sources.

Questions people ask about WebGPT

What is WebGPT?

It is a version of GPT-3 fine-tuned by OpenAI to answer long-form questions. It searches and navigates the web through a text-based browsing environment, and it collects references to back up what it says.

How was it trained?

Humans performed the browsing and answering task first, and the model learned from those demonstrations through imitation learning. Answer quality was then improved with human feedback, using a reward model trained to predict which answers people preferred.

How good were its answers?

On the ELI5 dataset, human evaluators preferred its best model's answers 56% of the time over those of the human demonstrators, and 69% of the time over the highest-voted Reddit answer for the same question.

Why does it cite sources?

The model is required to gather references while browsing so that people can more easily check whether the facts in an answer are accurate.

When was it published?

The paper was posted to arXiv on 17 December 2021, with later revisions in March and June 2022.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on WebGPT yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2112.09332, openai.com/index/webgpt/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.