Biasless Language Models Learn Unnaturally: How LLMs Fail to Distinguish the Possible from the Impossible

arXiv cs.CL / 4/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper investigates whether large language models (LLMs) can distinguish between humanly possible and impossible languages using learning-curve comparisons.
Replicating prior methodology across more languages and more types of “impossible” perturbations, the authors find GPT-2 typically learns natural languages and their impossible counterparts with similar ease.
A broader, more lenient test for separation between sets of possible vs. impossible languages also finds no systematic, consistent distinction in GPT-2 behavior.
Overall, the results suggest GPT-2 lacks a reliable bias or sensitivity to the possible/impossible distinction that has been hypothesized from earlier studies.

Abstract

Are large language models (LLMs) sensitive to the distinction between humanly possible and impossible languages? This question was recently used in a broader debate on whether LLMs and humans share the same innate learning biases. Previous work has answered it in the positive by comparing LLM learning curves on existing language datasets and on "impossible" datasets derived from them via various perturbation functions. Using the same methodology, we examine this claim on a wider set of languages and impossible perturbations. We find that in most cases, GPT-2 learns each language and its impossible counterpart equally easily, in contrast to previous findings. We also apply a more lenient condition by testing whether GPT-2 provides any kind of separation between the whole sets of natural vs. impossible languages, based on cross-linguistic variance in metrics derived from the learning curves. Taken together, these perspectives show that GPT-2 provides no systematic separation between the possible and the impossible.