CoNLL paper: transformers fail generation of 'impossible' English

Whether some grammars are simply unlearnable has usually been a claim about people. A new CoNLL 2026 paper from Ram Janarthan, Coleman Haley and Sharon Goldwater takes that question and asks it of transformer language models instead, and the answer is more interesting than a yes or no.

Earlier work in this area, the authors note, has mostly leaned on sample efficiency and test-set perplexity. Train a GPT-2 style model on an 'impossible' variant of English, watch it fit worse than on real English, and conclude that the model shares something with the human bias against those structures. Janarthan and colleagues argue that isn't enough evidence to link a model's behaviour to what actually rules certain grammars out of human languages. Perplexity tells you a model prefers one distribution to another; it does not tell you which linguistic capacity that preference reflects. So they test two candidate explanations directly: do the models fail to notice ungrammaticality, or do they fail to produce well-formed output?

The split they find is the payload. When grammatical sensitivity is probed with BLiMP minimal pairs, performance on the impossible variants drops only gradually, and the degradation tracks what they call the language's information locality. Generation is a different story. The same models show, in the authors' phrasing, 'pronounced failures in generation, producing substantially fewer high-quality sentences at longer lengths.' Impossibility is showing up as a production problem more than a comprehension one, which is a sharper claim than this literature has been making.

Scope matters here: these are GPT-2 scale models trained on perturbed English, and the abstract leaves open which perturbations were tested, whether frontier-scale models would show the same split, and how 'high-quality' output was scored, so treat the numbers as preliminary rather than final. The mechanism behind that split, why generation collapses when sensitivity does not, is left open by this write-up.

For anyone building linguistic evaluations of language models, the useful upshot is that 'got the perplexity right' and 'can actually use the grammar to produce coherent output' are not the same measurement, and the second one is where impossible languages seem to break.