When a problem becomes difficult, the model shifts to its dominant language — and becomes more precise

A study of eight languages shows four regimes of linguistic coherence, with a brutal collapse on non-Latin languages and the least represented.

A multilingual reasoning model is almost always judged on one thing: does it give the right answer? Muhammad Ali Shafique and Kelly Marchisio, in an article filed on August 9, 2026, ask another question — does the model reason and respond well in the expected language? Their analysis is based on the PolyMath testbed, which covers eight languages and four difficulty levels.

Four regimes emerge according to the difficulty: maintained alignment, persistent misalignment, progressive degradation and sudden collapse. The latter case is the most striking: beyond a certain threshold of difficulty, linguistic coherence drops sharply, and the effect strikes primarily non-Latin languages and those least represented in the training data.

The paradox identified by the authors is instructive for anyone who deploys these models outside of English: accuracy can improve at the most difficult levels at the very moment when the model switches to its dominant internal language. The study also examines the effect of quantification, where GPTQ and AWQ often outperform AutoRound, with effects on linguistic consistency and accuracy that are independent of each other.