modeLing: A Novel Dataset for Testing Linguistic Reasoning in Language Models

TL;DR

modeLing is a new benchmark of linguistics puzzles designed to evaluate AI systems' few-shot reasoning and compositional generalization, avoiding data leakage issues present in prior datasets.

Contribution

This work introduces modeLing, a novel, carefully crafted dataset of puzzles for testing linguistic reasoning in AI, ensuring no overlap with training data.

Findings

Large language models show some reasoning ability on modeLing

Models outperform random chance but still have significant errors

modeLing can serve as a benchmark for future improvements in linguistic reasoning

Abstract

We introduce modeLing, a novel benchmark of Linguistics Olympiad-style puzzles which tests few-shot reasoning in AI systems. Solving these puzzles necessitates inferring aspects of a language's grammatical structure from a small number of examples. Such puzzles provide a natural testbed for language models, as they require compositional generalization and few-shot inductive reasoning. Consisting solely of new puzzles written specifically for this work, modeLing has no risk of appearing in the training data of existing AI systems: this ameliorates the risk of data leakage, a potential confounder for many prior evaluations of reasoning. Evaluating several large open source language models and GPT on our benchmark, we observe non-negligible accuracy, demonstrating few-shot emergent reasoning ability which cannot merely be attributed to shallow memorization. However, imperfect model performance suggests that modeLing can be used to measure further progress in linguistic reasoning.