From Evidence to Belief: A Bayesian Epistemology Approach to Language Models
Abstract: This paper investigates the knowledge of language models from the perspective of Bayesian epistemology. Specifically, it aims to explore whether language models can accurately incorporate evidence of varying levels of informativeness and reliability into their confidence and responses. As Bayesian epistemology interprets belief as confidence according to evidence, this study offers a new perspective on understanding the beliefs and knowledge of language models. We created a dataset with various types of evidence and analyzed its response and confidence using verbalized confidence, token probability, and sampling. From the perspective of verbalized confidence, our research has shown that we can interpret that language models can generally reflect evidence in their confidence and calibration. We also demonstrated that language models exhibit biases toward correct evidence, exploit unreasonable evidence, and ignore errors in the context, all of which can be interpreted as the epistemic character of language models.
Paper Type: Long
Research Area: Interpretability and Analysis of Models for NLP
Research Area Keywords: probing, explanation, faithfulness, calibration
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 1263
Loading