Confidence Calibration in Large Language Models for Uncertainty Quantification: Affecting Calibration with Conditional Weight Updates

Abstract

In any medical applications of Large Language Models (LLMs), it is critical to have accurate uncertainty quantification, as well as control over the over- and under-confidence of the model. Current fine-tuning (FT) methods lack this control, partly because they fail to account for the fact that repeated exposure to a fact does not make it more correct. We propose a revised FT method that updates model weights only when the model does not sufficiently “know” an answer. We fine-tuned Meta's Llama-3.2, 1B parameter model on the MMLU multiple-choice dataset using traditional FT methods for a Control Model and Conditional Update FT for an Experi-mental Model. The tuned models showed different results, with the Control showing greater overconfidence and the Experimental Model showing greater under-confidence as compared to the Base Model. Additionally, the Experimental Model showed a more even distribution of confidence scores, which is advantageous for post-calibration. This method for affecting confidence calibration while fi-ne-tuning LLMs may potentially help in the broader challenge of creating reliable and trustworthy LLMs.