Major research into ‘hallucinating’ generative models advances reliability of artificial intelligence

With previous approaches, it wasn’t possible to tell the difference between a model being uncertain about what to say versus being uncertain about how to say it. But our new method overcomes this.

Study author Dr Sebastian Farquhar, Department of Computer Science, University of Oxford

The researchers focused on hallucinations where LLMs give different answers each time it is asked a question － even if the wording is identical － known as ‘confabulating’.

“LLMs are highly capable of saying the same thing in many different ways, which can make it difficult to tell when they are certain about an answer and when they are literally just making something up” said study author Dr Sebastian Farquhar, from the University of Oxford’s Department of Computer Science. “With previous approaches, it wasn’t possible to tell the difference between a model being uncertain about what to say versus being uncertain about how to say it. But our new method overcomes this.”

To do this, the research team developed a method grounded in statistics and using methods that estimate uncertainty based on the amount of variation (measured as entropy) between multiple outputs. Their approach computes uncertainty at the level of meaning rather than sequences of words, i.e. it spots when LLMs are uncertain about the actual meaning of an answer, not just the phrasing. To do this, the probabilities produced by the LLMs, which state how likely each word is to be next in a sentence, are translated into probabilities over meanings.

Our method basically estimates probabilities in meaning-space, or ‘semantic probabilities. The appeal of this approach is that it uses the LLMs themselves to do this conversion.

Study co-author Jannik Kossen, Department of Computer Science, University of Oxford

The new method proved much better at spotting when a question was likely to be answered incorrectly than all previous methods, when tested against six open-source LLMs (including GPT-4 and LLaMA 2). This was the case for a wide range of different datasets including answering questions drawn from Google searches, technical biomedical questions, and mathematical word problems. The researchers even demonstrated how semantic entropy can identify specific claims in short biographies generated by ChatGPT that are likely to be incorrect.

“Our method basically estimates probabilities in meaning-space, or ‘semantic probabilities’,” said study co-author Jannik Kossen (Department of Computer Science, University of Oxford). “The appeal of this approach is that it uses the LLMs themselves to do this conversion.”

By detecting when a prompt is likely to produce a confabulation, the new method can help make users of generative AI aware when the answers to a question are probably unreliable, and to allow systems built on LLMs to avoid answering questions likely to cause confabulations. A key advantage to the technique is that it works across datasets and tasks without a priori knowledge, requiring no task-specific data, and robustly generalises to new tasks not seen before. Although it can make the process several times more computationally costly than just using a generative model directly, this is clearly justified when accuracy is paramount.

Making large language models reliable and trustworthy remains a huge challenge and opportunity.

 Dr Sebastian Farquhar, Department of Computer Science, University of Oxford

Currently, hallucinations are a critical factor holding back wider adoption of LLMs like ChatGPT or Gemini. Besides making LLMs unreliable, for example by presenting inaccuracies in news articles and fabricating legal precedents, they can even be dangerous, for example when used in medical diagnosis.

The study’s senior author Yarin Gal, Professor of Computer Science at the University of Oxford and Director of Research at the UK’s AI Safety Institute, said: “Getting answers from LLMs is cheap, but reliability is the biggest bottleneck. In situations where reliability matters, computing semantic uncertainty is a small price to pay.”

Professor Gal’s research group, the Oxford Applied and Theoretical Machine Learning group, is home to this and other work pushing the frontiers of robust and reliable generative models. Building on this expertise, Professor Gal now acts as Director of Research at the UK’s AI Safety Institute.

The researchers highlight that confabulation is just one type of error that LLMs can make. “Semantic uncertainty helps with specific reliability problems, but this is only part of the story,” explained Dr Farquhar. “If an LLM makes consistent mistakes, this new method won’t catch that. The most dangerous failures of AI come when a system does something bad but is confident and systematic. There is still a lot of work to do.”

The study ‘Detecting Hallucinations in Large Language Models Using Semantic Entropy’ has been published in Nature. The study was partially funded by the Alan Turing Institute.