Large Language Model Confidence Estimation via Black-Box Access

Given the proliferation of deep learning over the last decade or so [5], uncertainty or confidence estimation of these models has been an active research area [4]. Predicting accurate confidences in the generations produced by a large language model (LLM) are crucial for eliciting trust in the model and is also helpful for benchmarking and ranking competing models [37]. Moreover, LLM hallucination detection and mitigation, which is one of the most pressing problems in artificial intelligence research today [33], can also benefit significantly from accurate confidence estimation as it would serve as a strong indicator of the faithfulness of a LLM response. This applies to even settings where strategies such as retrieval augmented generation (RAG) are used [3] to mitigate hallucinations. Methods for confidence estimation in LLMs assuming just black-box or query access have been explored only recently [14, 19] and this area of research is still largely in its infancy. However, effective solutions here could have significant impact given their low requirement (i.e.