“WebGPT: Improving the Factual Accuracy of Language Models through Web Browsing”, 2021-12-16 (; similar):
[paper; prompt-based reimplementation, OA API] We’ve fine-tuned GPT-3 to more accurately answer open-ended questions using a text-based web browser. Our prototype copies how humans research answers to questions online—it submits search queries, follows links, and scrolls up and down web pages. It is trained to cite its sources, which makes it easier to give feedback to improve factual accuracy. We’re excited about developing more truthful AI, but challenges remain, such as coping with unfamiliar types of questions.
Language models like GPT-3 are useful for many different tasks, but have a tendency to “hallucinate” information when performing tasks requiring obscure real-world knowledge.23 To address this, we taught GPT-3 to use a text-based web-browser. The model is provided with an open-ended question and a summary of the browser state, and must issue commands such as “Search …”, “Find in page: …” or “Quote: …”. In this way, the model collects passages from web pages, and then uses these to compose an answer.
The model is fine-tuned from GPT-3 using the same general methods we’ve used previously. We begin by training the model to copy human demonstrations, which gives it the ability to use the text-based browser to answer questions. Then we improve the helpfulness and accuracy of the model’s answers, by training a reward model to predict human preferences, and optimizing against it using either reinforcement learning or rejection sampling.
…Our models outperform GPT-3 on TruthfulQA and exhibit more favourable scaling properties. However, our models lag behind human performance, partly because they sometimes quote from unreliable sources (as shown in the question about ghosts above). We hope to reduce the frequency of these failures using techniques like adversarial training.
…Evaluating factual accuracy: …
However, this approach raises a number of questions. What makes a source reliable? What claims are obvious enough to not require support? What trade-off should be made between evaluations of factual accuracy and other criteria such as coherence? All of these were difficult judgment calls. We do not think that our model picked up on much of this nuance, since it still makes basic errors. But we expect these kinds of decisions to become more important as AI systems improve, and cross-disciplinary research is needed to develop criteria that are both practical and epistemically sound. We also expect further considerations such as transparency to be important.
Eventually, having models cite their sources will not be enough to evaluate factual accuracy. A sufficiently capable model would cherry-pick sources it expects humans to find convincing, even if they do not reflect a fair assessment of the evidence. There are already signs of this happening (see the questions about boats above). We hope to mitigate this using methods like debate.