LLMs as Signal Detectors: Sensitivity, Bias, and the Temperature-Criterion Analogy

Large language models' calibration is typically assessed using metrics like Expected Calibration Error, which fails to distinguish between two crucial aspects: sensitivity and bias. Signal Detection Theory provides a framework to decompose these components, allowing for a more nuanced evaluation of LLMs. By applying SDT-derived metrics, such as the area under the receiver operating characteristic curve, researchers can separately examine a model's ability to discern correct from incorrect answers and its propensity for confident or cautious responses¹. This distinction is critical, as it enables the identification of models that are overly confident or prone to false positives. The temperature-criterion analogy further refines this analysis, offering insights into the interplay between sensitivity and bias. This matters to practitioners, as a deeper understanding of LLMs' strengths and weaknesses is essential for developing reliable and trustworthy AI systems.