Black-box LLM uncertainty splits in two: a two-level ensemble separates vague prompts from missing knowledge
Synopsis
Large language model uncertainty has been collapsed into a single consistency score that cannot say whether the model misread the question or lacks the answer. A two-level ensemble treats repeated stochastic decodes of one prompt as the inner level and meaning-preserving rewrites as the outer level, then uses the law of total covariance in sentence-embedding space to split aleatoric from epistemic uncertainty using only black-box sampling. Across five short QA benchmarks and four instruction-tuned models it is competitive with strong white-box baselines, and paired interventions make the two components respond separately to prompt ambiguity and knowledge gaps.
Interpretation
A black-box LLM call can now return two kinds of uncertainty, one from ambiguity in the prompt and one from gaps in the model's knowledge, instead of a single generic consistency score. Earlier black-box methods collapsed output variability into one scalar that flagged unreliability without diagnosing its source; Bayesian methods and deep ensembles can in principle make the split but are computationally prohibitive at billions of parameters and need internal access. On five short QA benchmarks and four instruction-tuned models spanning the LLaMA-3 and Qwen2.5 families at two scales each, the estimator is competitive with strong white-box baselines; at a matched generation budget it matches leading one-level black-box methods on error ranking while also decomposing the two uncertainties.
The method samples at two levels: repeated stochastic decodes of the same prompt form the inner level, and meaning-preserving rewrites form the outer level, with output spread measured in sentence-embedding space. The law of total covariance splits total variability exactly into within-perturbation spread, which maps to aleatoric uncertainty, and disagreement of conditional means across perturbations, which maps to epistemic uncertainty; the inner sample mean carries a bias that a closed-form correction removes. Outer rewrites are drawn by snapping embedding-space Gaussian noise to a cached paraphrase pool, and inner decodes run at fixed deployed parameters; on AmbigQA a gold clarification significantly shifts aleatoric uncertainty, on Natural Questions gold evidence significantly shifts epistemic uncertainty, and fixed-prompt consistency baselines stay statistically unchanged after clarification.
Perspective
The demonstrated scope is short-form QA at fixed model weights: five short QA benchmarks and four instruction-tuned models, plus a feasibility pilot that scores atomic claims on the Natural Questions long-answer split; the paired interventions target one uncertainty source each on AmbigQA and Natural Questions, and the score-level comparison is mainly on LLaMA-3-8B.
The estimator targets perturbation-population uncertainty at fixed weights, and its correspondence to Bayesian parameter uncertainty relies on local linearity, a locally Gaussian posterior and semantic coverage, which are measured as diagnostics rather than guaranteed; automatic paraphrasing can introduce semantic drift that inflates between-perturbation variance; uncertainty estimates inherit the sentence encoder's representational geometry and blind spots; nested sampling needs more decodes per query than one-level baselines; and the long-form pilot has no human verification and does not test judge bias against the generator family.
