GPT-2 models fail to generate 'impossible' human languages, study shows

A new paper from researchers at the University of Edinburgh (Ram Janarthan, Coleman Haley, and Sharon Goldwater) investigates why transformer language models prefer human languages over 'impossible' ones that humans cannot acquire. Using GPT-2 style models trained on artificially perturbed English (e.g., scrambled word order or non-local dependencies), they evaluated two hypotheses: that impossibility stems from poor grammatical sensitivity or from poor generative ability. The study used BLiMP minimal pairs to measure grammaticality sensitivity and analyzed sentence generation quality.

The results show that model performance on grammaticality judgments only gradually degrades, with the severity mediated by the language's information locality—a measure of how locally information is structured. In contrast, generative output showed dramatic failures: models produced substantially fewer high-quality sentences, especially at longer lengths. This asymmetry points to generative deficiency and transmission breakdown as the key explanation for why transformers fail on unnatural languages, rather than an inability to learn grammatical patterns. The paper won the Best Paper Award at CoNLL 2026, highlighting its significance in understanding the link between LM behavior and human language acquisition.