A Unified Assessment of the Poverty of the Stimulus Argument for Neural Language Models

Xiulin Yang, Arianna Bisazza, Nathan Schneider, Ethan Gotlieb Wilcox · 2026

Read A Unified Assessment of the Poverty of the Stimulus Argument for Neural Language Models by Xiulin Yang, Arianna Bisazza, Nathan Schneider, Ethan Gotlieb Wilcox (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-2602-09992.txt.

Description

Linguists have long debated whether children could learn grammar from the limited speech they hear without built-in syntactic knowledge. A 2026 paper led by Xiulin Yang, with three co-authors, tests that poverty-of-the-stimulus question for neural language models. The team builds a unified benchmark spanning four canonical phenomena and trains Transformer, LSTM, and n-gram models under shared protocols.

The models generalize above chance from as little as ten million words, yet they learn less efficiently than children as input grows. Cognitively motivated biases improve broad syntactic skill but do not reliably help on the stimulus-poor phenomena. The findings push back on the idea that built-in syntax is the sole route to generalization, while suggesting human-like efficiency still needs biases beyond those tested.

Catalog record

CREATORXiulin Yang, Arianna Bisazza, Nathan Schneider, Ethan Gotlieb Wilcox
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-09992

Browse the full library · About the Tower of Babel