SemanticDPP: Efficient Uncertainty Quantification in LLMs

Jenny Hamer, Jasper Snoek, Alexander Nicholas D'Amour, Ulrich Rueckert, Dheeraj Rajagopal, Victor Veitch, Jie Ren, Kevin Robinson

Keywords: uncertainty, semantic, large language models, determinantal point processes

TL;DR: We efficiently quantify large language model uncertainty, using determinantal point processes (DPPs), by learning a semantically meaningful subspace of model embeddings and clustering sampled model responses in this space.

Abstract:

Accurately quantifying uncertainty is crucial for ensuring the reliability of model outputs and enabling informed downstream decision-making. However, the output space of large language models (LLMs) is so large that traditional methods break down in this setting; yet, LLMs have been shown to respond to prompts confidently with confabulations, fabrications, ``hallucinations’’, and other erroneous information. While methods designed specifically for LLMs exist, there is a gap and need for approaches which accurately quantify uncertainty efficiently. In this work, we introduce SemanticDPP, an efficient method for quantifying the uncertainty corresponding to the semantic variation in outputs of a large language model based on internal activations. Building upon prior work, we show that this uncertainty is highly indicative of whether or not a model will answer correctly in question answering (QA) tasks---which provides a useful signal to determine whether to abstain from responding to a given question to prevent incorrect answers.
SemanticDPP uses determinantal point processes (DPPs) to learn a model over semantic (dis)similarity in model embeddings, which we use in a fully unsupervised manner to identify semantically distinct sets of responses. We additionally present an extension, SemanticDPP-C, which yields a soft clustering of the sets of semantically distinct responses. Our extensive empirical investigation examines the behavior of our methods on two frontier open-sourced models of different capacities (that grant access to model internals), Gemma2 9B and Gemma3 27B, on a broad range of widely used QA benchmarks. SemanticDPP enables fast uncertainty quantification while it matches or exceeds the selective prediction (hallucination detection) performance of state-of-the-art baselines.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

Submission Number: 9298