Quantifying Aleatoric Uncertainty of In-Context Learning for Robust Measure of LLM Prediction Confidence

The paper introduces a novel approach to quantify aleatoric uncertainty in In-Context Learning (ICL) for large language models (LLMs) by utilizing self-function vectors, which are based on Bayesian principles and mechanistic interpretability. This method allows for a direct estimation of aleatoric uncertainty, distinguishing it from epistemic uncertainty, and is supported by a new rigorous evaluation protocol that manipulates data in controlled ways. The findings suggest that this approach enhances the reliability of LLM predictions under ICL, with potential applications in areas like hallucination detection, thereby improving trustworthiness in AI systems.