Large language models generate text by sampling from a probability distribution over possible tokens. That distribution carries an internal estimate of how likely each response is. Separately, models can be prompted to state their confidence in plain language, such as "I am 80% sure."
According to a new preprint on arXiv, these two forms of uncertainty are not independent. The paper's title and abstract indicate that verbalized confidence is coupled with the internal probabilities the model assigns to its outputs. This suggests that when a model says it is uncertain, that statement may reflect something real about its underlying computations.
If the coupling holds across models and tasks, it could make verbalized confidence a practical tool for calibration. Users might rely on a model's stated uncertainty as a signal for when to trust its output. The preprint is a single source, so the findings await peer review and replication, but the result points toward a more interpretable link between a model's words and its internal states.