Epistemic Uncertainty for Test-Time Discovery

Kainat Riaz, Muhammad Ahmed Mohsin, Ahsan Bilal, Muhammad Umer, Ayesha Mohsin, Aqib Riaz, Ali Subhan, John M. Cioffi, [0.5em] Stanford University · 2026

Read Epistemic Uncertainty for Test-Time Discovery by Kainat Riaz, Muhammad Ahmed Mohsin, Ahsan Bilal, Muhammad Umer, Ayesha Mohsin, Aqib Riaz, Ali Subhan, John M. Cioffi, [0.5em] Stanford University (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-2605-11328.txt.

Description

Riaz and seven colleagues tackle test-time scientific discovery, where standard reinforcement learning penalizes high-variance mutations and caps the best reward. Their UG-TTT method keeps a small set of low-rank adapters laid over a frozen base. Disagreement across adapters, measured as mutual information, marks epistemic uncertainty, or gaps in coverage, rather than hard problems. Added as an exploration bonus, it raises the maximum reward on three of four benchmarks while keeping solution diversity high.

Catalog record

CREATORKainat Riaz, Muhammad Ahmed Mohsin, Ahsan Bilal, Muhammad Umer, Ayesha Mohsin, Aqib Riaz, Ali Subhan, John M. Cioffi, [0.5em] Stanford University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64330
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.11328

Browse the full library · About the Tower of Babel