Heuristic Pathologies and Further Variance Reduction via Uncertainty Propagation in the AIVAT Family of Techniques

Juho Kim, Tuomas Sandholm · 2026

Read Heuristic Pathologies and Further Variance Reduction via Uncertainty Propagation in the AIVAT Family of Techniques by Juho Kim, Tuomas Sandholm (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-2605-14261.txt.

Description

Kim and Sandholm expose heuristic flaws in AIVAT variance reduction and spread heuristic uncertainty. AIVAT rates agent results in multiagent settings with few samples or costly trials, using unbiased low-variance estimates. A heuristic value function splits low- and high-value counterfactuals, yet little guidance exists on choosing it. Sample variance can be gamed low by gradient descent, and desired results can be p-hacked via gradient moves. Fix the function before seeing the evaluation data. That uncertainty quantifies AIVAT estimates, and averaging cuts variance at the cost of unbiasedness.

Catalog record

CREATORJuho Kim, Tuomas Sandholm
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE55943
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.14261

Browse the full library · About the Tower of Babel