LLM Confidence Reflects Latent Decision Variables

For large language models (LLMs) to be trustworthy, their confidence in an answer must reliably reflect its correctness. While existing work often evaluates confidence by its predictive power or calibration, a deeper question remains: what does the confidence signal itself truly represent? Is it a normative reflection of a latent decision variable, or merely a heuristic preference?This study addresses this by applying Statistical Decision Confidence (SDC), a framework from computational neuroscience. By treating the answer-logit difference (LD) as a potential readout of a latent decision variable, researchers tested SDC's qualitative signatures. Across several perceptual discrimination and memory tasks, involving both multimodal non-reasoning and reasoning models, LD consistently satisfied these signatures. This indicates that, in these contexts, answer logits function as monotonic readouts of a latent decision variable, rather than simple heuristic scores.However, in more complex visual reasoning tasks, while LD still predicted correctness, the full geometric signatures of SDC were absent. This highlights the current limitations of the framework when explicit normative process models are unavailable. The findings offer a computational explanation for confidence in multimodal language models and establish SDC as a unifying framework for studying confidence across biological and artificial intelligence.