Researchers say a single confidence score tells you almost nothing useful about why an AI model got something wrong.
A new paper proposes splitting LLM uncertainty into three separate problems: ambiguous input prompts, gaps in the model's stored knowledge, and randomness introduced during text generation. The authors argue that the traditional split between "aleatoric" and "epistemic" uncertainty — borrowed from classical statistics — is too coarse to drive practical improvements. Through experiments across model sizes and task types, they show that which source of uncertainty dominates can shift dramatically depending on context, meaning a fix that works for one failure mode may do nothing for another.
The distinction matters because the remedies are different. Prompt ambiguity is a user or system design problem. Knowledge gaps point to training data or retrieval augmentation. Decoding randomness is a sampling parameter question. Lumping them into one score leaves engineers guessing at the wrong lever — and leaves hallucination detection blunt when it needs to be precise.
Most commercial LLM reliability tooling still reports a single confidence figure, which this framework implicitly indicts as insufficient. Whether the decomposition holds up outside controlled experiments — and whether it scales to production inference pipelines — remains to be seen.