🛰️ Daily AI Frontier
‹ back to 2026-07-16

Subjective Risk Decomposition: A New View for Uncertainty Quantification

Research LLMs & Foundation Models

Merged summary

TL;DR - A preprint proposing that uncertainty measures aren't axiomatic primitives but consequences of modelling choices, deriving epistemic and aleatoric uncertainty by decomposing a "subjective risk" built from a strictly proper loss. It matters because it unifies many scattered UQ measures under one theoretical foundation and links them to learning theory.

  • Reframes uncertainty quantification (UQ): given a modelling scenario and a strictly proper loss, the epistemic/aleatoric terms are induced by decomposing the subjective risk rather than postulated.
  • Reverse cross-entropy is a headline example, where the decomposition recovers the classic information-theoretic uncertainty terms.
  • Claims the same approach reproduces numerous previously proposed UQ measures, giving them a common grounding.
  • Extends to learning theory via subjective-risk analogues of excess risk, approximation error, and estimation error, positioned as a first step toward a learning-theoretic UQ framework.

Note: summary is based solely on the abstract; no empirical results or benchmarks are provided in the content.

Sources (1)

Subjective Risk Decomposition: A New View for Uncertainty Quantification

arXiv stat.ML Raghad Alamri, Michele Caprio, Gavin Brown 2026-07-16 arXiv:2607.15196

TL;DR - A preprint proposing that uncertainty measures aren't axiomatic primitives but consequences of modelling choices, deriving epistemic and aleatoric uncertainty by decomposing a "subjective risk" built from a strictly proper loss. It matters because it unifies many scattered UQ measures under one theoretical foundation and links them to learning theory.

  • Reframes uncertainty quantification (UQ): given a modelling scenario and a strictly proper loss, the epistemic/aleatoric terms are induced by decomposing the subjective risk rather than postulated.
  • Reverse cross-entropy is a headline example, where the decomposition recovers the classic information-theoretic uncertainty terms.
  • Claims the same approach reproduces numerous previously proposed UQ measures, giving them a common grounding.
  • Extends to learning theory via subjective-risk analogues of excess risk, approximation error, and estimation error, positioned as a first step toward a learning-theoretic UQ framework.

Note: summary is based solely on the abstract; no empirical results or benchmarks are provided in the content.

item →