Factual Confidence of LLMs: on Reliability and Robustness of Current Estimators
This work addresses the problem of unreliable factuality in LLMs for NLP researchers and practitioners, providing a benchmark for confidence estimation methods, but it is incremental as it surveys and compares existing techniques rather than introducing new ones.
The paper systematically compares methods for estimating factual confidence in Large Language Models (LLMs) and finds that trained hidden-state probes are most reliable, though they require access to model weights and training data, while also revealing that LLM confidence is often unstable across semantically equivalent inputs.
Large Language Models (LLMs) tend to be unreliable in the factuality of their answers. To address this problem, NLP researchers have proposed a range of techniques to estimate LLM's confidence over facts. However, due to the lack of a systematic comparison, it is not clear how the different methods compare to one another. To fill this gap, we present a survey and empirical comparison of estimators of factual confidence. We define an experimental framework allowing for fair comparison, covering both fact-verification and question answering. Our experiments across a series of LLMs indicate that trained hidden-state probes provide the most reliable confidence estimates, albeit at the expense of requiring access to weights and training data. We also conduct a deeper assessment of factual confidence by measuring the consistency of model behavior under meaning-preserving variations in the input. We find that the confidence of LLMs is often unstable across semantically equivalent inputs, suggesting that there is much room for improvement of the stability of models' parametric knowledge. Our code is available at (https://github.com/amazon-science/factual-confidence-of-llms).