Credal Concept Bottleneck Models for Epistemic–Aleatoric Uncertainty Decomposition

Tanmoy Mukherjee, Thomas Bailleux, Pierre Marquis, Zied Bouraoui · 2026

Read Credal Concept Bottleneck Models for Epistemic–Aleatoric Uncertainty Decomposition by Tanmoy Mukherjee, Thomas Bailleux, Pierre Marquis, Zied Bouraoui (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-24170.txt.

Description

Mukherjee and three colleagues introduce CREDENCE, a concept bottleneck framework that splits concept uncertainty into epistemic and aleatoric parts. Standard bottleneck models output single concept chances that blur model gaps with input ambiguity that cannot shrink. CREDENCE gives each concept a range, reads epistemic uncertainty from disagreement across concept heads, and estimates aleatoric uncertainty from an ambiguity output trained on rater disagreement. The split supports clear actions: automate easy cases, gather data where knowledge is thin, send unclear cases to humans, and abstain when both run high.

Catalog record

CREATORTanmoy Mukherjee, Thomas Bailleux, Pierre Marquis, Zied Bouraoui
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE72079
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.24170

Browse the full library · About the Tower of Babel