Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs
Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, Yarin Gal
2024-06-25
Summary
This paper introduces Semantic Entropy Probes (SEPs), a new method for detecting hallucinations in large language models (LLMs). Hallucinations are when the model generates responses that sound plausible but are actually incorrect or nonsensical.
What's the problem?
Hallucinations in LLMs can be a major issue because they undermine the reliability of these models in real-world applications. Traditional methods for detecting these hallucinations often require a lot of computational resources, making them impractical for widespread use.
What's the solution?
The authors propose SEPs as a more efficient way to detect hallucinations. Instead of needing to generate multiple outputs to check for accuracy, SEPs work by analyzing the hidden states of a single model generation. This approach allows for quick and reliable detection of uncertainty in the model's responses without the heavy computational costs associated with other methods. The researchers demonstrate that SEPs perform well in identifying hallucinations and can adapt to new tasks effectively.
Why it matters?
This research is important because it offers a cost-effective solution for improving the reliability of LLMs. By making it easier to detect hallucinations, SEPs can enhance the practical use of these models in various applications, such as chatbots and automated content generation, ensuring that they provide more accurate and trustworthy information.
Abstract
We propose semantic entropy probes (SEPs), a cheap and reliable method for uncertainty quantification in Large Language Models (LLMs). Hallucinations, which are plausible-sounding but factually incorrect and arbitrary model generations, present a major challenge to the practical adoption of LLMs. Recent work by Farquhar et al. (2024) proposes semantic entropy (SE), which can detect hallucinations by estimating uncertainty in the space semantic meaning for a set of model generations. However, the 5-to-10-fold increase in computation cost associated with SE computation hinders practical adoption. To address this, we propose SEPs, which directly approximate SE from the hidden states of a single generation. SEPs are simple to train and do not require sampling multiple model generations at test time, reducing the overhead of semantic uncertainty quantification to almost zero. We show that SEPs retain high performance for hallucination detection and generalize better to out-of-distribution data than previous probing methods that directly predict model accuracy. Our results across models and tasks suggest that model hidden states capture SE, and our ablation studies give further insights into the token positions and model layers for which this is the case.