Directional Concentration Uncertainty: A representational approach to uncertainty quantification for generative models

Souradeep Chattopadhyay, Brendan Kennedy, Sai Munikoti, Soumik Sarkar, Karl Pazdernik · 2026

Read Directional Concentration Uncertainty: A representational approach to uncertainty quantification for generative models by Souradeep Chattopadhyay, Brendan Kennedy, Sai Munikoti, Soumik Sarkar, Karl Pazdernik (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-2602-13264.txt.

Description

Trust in generative models depends on knowing when an output is shaky, yet many uncertainty tools lean on narrow rules that fail on new tasks or media. This paper proposes Directional Concentration Uncertainty, a method that reads the spread of a model's many outputs in embedding space through a directional statistics lens. When outputs cluster tightly in one direction, confidence is high; when they scatter, it is low.

The approach needs no task-specific tricks, and tests show it matches or beats prior tools such as semantic entropy on calibration, while also carrying over to multi-modal tasks. That breadth makes it a candidate for safety checks inside agentic systems, where a model must judge its own doubt before acting. Souradeep Chattopadhyay leads the work with four co-authors.

Catalog record

CREATORSouradeep Chattopadhyay, Brendan Kennedy, Sai Munikoti, Soumik Sarkar, Karl Pazdernik
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE31 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-13264

Browse the full library · About the Tower of Babel