Deciphering Shortcut Learning from an Evolutionary Game Theory Perspective

Xiayang Li [email protected], Kuo Gai [email protected], Shihua Zhang [email protected] · 2026

Read Deciphering Shortcut Learning from an Evolutionary Game Theory Perspective by Xiayang Li [email protected], Kuo Gai [email protected], Shihua Zhang [email protected] (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-2605-02658.txt.

Description

Xiayang Li, Kuo Gai, and Shihua Zhang explain, with evolutionary game theory, why deep networks learn shortcuts. Shortcut learning means a model leans on easy, non-essential traits in the data instead of the real signal. Its roots in training dynamics lacked a formal account, which this paper supplies by defining core and shortcut features precisely.

Data samples are cast as players, and their neural tangent features as strategies, with core and shortcut subnetworks assumed to exist. Plain gradient descent and its stochastic form settle into different stable states: the first mainly trains the shortcut subnetwork, the second the core one. A continuous stochastic equation then traces how data noise and optimization noise shape shortcut bias. The theory offers a principled view of how to reduce the bias.

Catalog record

CREATORXiayang Li [email protected], Kuo Gai [email protected], Shihua Zhang [email protected]
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE128 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.02658

Browse the full library · About the Tower of Babel