Deploying large language models (LLMs) reliably requires distinguishing uncertainty that stems from irreducible task variability (aleatoric) from uncertainty caused by gaps in the model’s knowledge (epistemic). In language tasks, input ambiguity or underspecification often yields several plausible interpretations, which is a primary source of aleatoric uncertainty.
Current decomposition approaches typically generate multiple clarifications of the input, query the model for an answer under each clarification, and then compare the answers to estimate aleatoric uncertainty. We argue that this answer‑centric step is unnecessary: answers are frequently redundant, increase computational cost, and can leak epistemic information that contaminates the aleatoric estimate.
Theoretically, we show that the space of plausible interpretations alone suffices to quantify ambiguity‑induced aleatoric uncertainty, without requiring model outputs for each clarification. Building on this insight, we propose a clarification‑only estimator: first detect a set of reasonable interpretations, then compute a divergence measure over this interpretation distribution to obtain an ambiguity score.
Using ambiguity detection as an operational evaluation across three benchmarks, the direct approach achieves an AUROC of 63.34% (versus 60.85% for the answer‑based baseline), reduces output tokens by 4‑26×, cuts API calls by 2.2‑3.5×, and yields uncertainty estimates with markedly lower correlation to epistemic uncertainty.
Overall, the results suggest that estimating ambiguity‑induced aleatoric uncertainty from the interpretation space is more accurate and efficient than from the response space.
Review