Improving Metacognition and Uncertainty Communication in Language Models (2510.05126v1)

Abstract: LLMs are increasingly used in decision-making contexts, but when they present answers without signaling low confidence, users may unknowingly act on erroneous outputs. While prior work shows that LLMs maintain internal uncertainty signals, their explicit verbalized confidence is typically miscalibrated and poorly discriminates between correct and incorrect answers. Across two types of LLMs, we investigate whether supervised finetuning can improve models' ability to communicate uncertainty and whether such improvements generalize across tasks and domains. We finetune the LLMs on datasets spanning general knowledge, mathematics, and open-ended trivia, and evaluate two metacognitive tasks: (1) single-question confidence estimation, where the model assigns a numeric certainty to its answer, and (2) pairwise confidence comparison, where the model selects which of two answers it is more likely to have correct. We assess generalization to unseen domains, including medical and legal reasoning. Results show that finetuning improves calibration (alignment between stated confidence and accuracy) and discrimination (higher confidence for correct vs. incorrect responses) within and across domains, while leaving accuracy unchanged. However, improvements are task-specific: training on single-question calibration does not transfer to pairwise comparison, and vice versa. In contrast, multitask finetuning on both forms of metacognition yields broader gains, producing lower calibration error and stronger discrimination in out-of-domain evaluations. These results show that while uncertainty communication in LLMs is trainable and generalizable, different metacognitive skills do not naturally reinforce one another and must be developed together through multitask training.