An Uncorrelated Pair Is Not a Separated Source: What a Disentanglement Score Certifies About Aleatoric and Epistemic Uncertainty, Why Correlation, Proxy Tasks and Ground-Truth Targets Rank the Same Methods Differently, and the Decision That Has to Be Declared Before Any of Them Counts

Abstract

Life of Research

Separating a model's uncertainty into an aleatoric part, attributed to noise in the task, and an epistemic part, attributed to what the model has not learned, is now a standard goal, and the 2024-2026 literature has returned three different verdicts on whether current methods achieve it. A 2024 benchmark on ImageNet reported that the aleatoric and epistemic estimates produced by the information-theoretic decomposition are rank-correlated between 0.78 and 0.99 for every distributional method it tested, and read this as a failure to disentangle. A 2026 semi-synthetic benchmark with known posteriors finds that the true components are themselves correlated (0.42 between posterior variance and aleatoric noise) and that, measured against posterior variance, excess entanglement is smaller than reported, while all or nearly all of its 17 methods track posterior variance more closely than the posterior risk it proposes as ground truth. A second 2026 study argues that correlation is the wrong test altogether: on CIFAR-10H the true components correlate at up to 0.882, and an evidential network whose estimates correlate at 0.999 achieves the best regret-based score. In this analysis the author reviews and synthesises published results to show that these verdicts do not conflict once the referent each one scores against is named. "Disentangled" is used for five different properties - success on a proxy task, low correlation between estimates, agreement with posterior spread, agreement with posterior excess risk, and sufficiency for a declared abstention decision - and the published evidence shows that they rank the same methods differently. The paper consolidates 35 published results in a table and a figure, adds two complications to the popular picture (the aleatoric estimate absorbs epistemic error, and the information-theoretic split has a finite-sample geometry that couples the components), gives a procedure for reading a disentanglement claim, and specifies the studies that would settle which referent a given use requires.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 144 papers by Pranay Mahendrakar, or his ORCID record.