Studies with impossible languages falsify LMs as models of human language

TL;DR

This paper reviews evidence showing that language models often learn impossible languages as well as attested ones, highlighting differences from human language acquisition and suggesting missing human-like biases in LMs.

Contribution

It demonstrates that LMs do not replicate human learning biases, as they learn impossible languages with similar ease as natural ones, challenging assumptions about their cognitive modeling.

Findings

LMs often learn impossible languages as well as natural languages

Impossible languages tend to be more complex or random

LMs lack human inductive biases for language learning

Abstract

According to Futrell and Mahowald [arXiv:2501.17047], both infants and language models (LMs) find attested languages easier to learn than impossible languages that have unnatural structures. We review the literature and show that LMs often learn attested and many impossible languages equally well. Difficult to learn impossible languages are simply more complex (or random). LMs are missing human inductive biases that support language acquisition.