Verbalized Confidence Across Autoregressive and Diffusion Language Models

We evaluate whether verbalized confidence is comparable and calibrated across autoregressive language models (AR models) and diffusion language models (DLMs). The evaluation includes three AR models, four DLMs, 250 questions from five datasets, and three prompt conditions. Each model returns an answer, confidence score, and short explanation under a shared output schema, and we evaluate correctness, calibration, selective prediction, and high-confidence error. Under the neutral prompt, AR models are better calibrated in aggregate, but are diverse in capability: Mercury-2 is competitive with strong AR models, while Dream and LLaDA show low accuracy and expected calibration error above 0.55. Prompting pressure worsens calibration for both model families. These results suggest that verbalized confidence should be treated as a behavioral variable depending on (but not limited to) model architecture, prompting condition, dataset and answer type. The conclusions are preliminary and requires scaled data, control for task type and difficulty, manual judge validation, and internal generation signals.