arXiv paper proposes black-box method to split LLM uncertainty into ambiguity and knowledge gaps
A two-level ensemble estimates aleatoric and epistemic uncertainty from black-box sampling alone, with results reported on five short QA benchmarks and four instruction-tuned models.
In plain words
When a language model answers, it can be unclear whether the question itself is ambiguous or whether the model simply lacks the knowledge. The paper describes a way to separate these two cases using only the model's outputs, without access to its internal workings. It does this by sampling many answers and measuring how variability changes when the prompt is reworded without changing its meaning.
What was announced
The authors, in an arXiv abstract, argue that reliable use of large language models needs uncertainty estimates that distinguish ambiguity inherent in a prompt from genuine gaps in the model's knowledge. They note that Bayesian and deep-ensemble approaches can make this aleatoric/epistemic split in principle but are computationally prohibitive at LLM scale and require white-box access, while existing black-box methods collapse uncertainty into a single scalar that cannot draw the distinction.
Their method builds a two-level ensemble: stochastic decoding of a fixed prompt forms the inner ensemble, and meaning-preserving perturbations of the prompt form the outer one. Variability is measured in a continuous embedding space, and the law of total covariance yields an exact within-/between-perturbation decomposition of aleatoric and epistemic uncertainty, requiring only black-box sampling access. The authors report three theoretical results, including a bias-corrected epistemic estimator and a separation result showing failure modes that consistency-only black-box methods provably miss. On five short QA benchmarks and four instruction-tuned LLMs across two families and two scales, they say the estimator is competitive with strong white-box baselines, matches leading black-box methods on error ranking at a matched generation budget, and a long-form pilot on the NQ long-answer split suggests extension to claim-level factuality without retraining.
What it does not show
- The evidence is the paper's own abstract; results are author-reported and not independently confirmed, and the abstract does not include the full article's details.
- Reported evaluation covers five short QA benchmarks and four instruction-tuned LLMs; the long-form result is described only as a pilot on one NQ split.
For security and IT teams
- Scope
- The method requires only black-box sampling access to the model, not white-box access to internal weights or activations, according to the abstract.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.