We present experimental evidence that English—the dominant language in large language model training—actively interferes with learning structurally explicit patterns. Through controlled experiments comparing (1) French vs. English, (2) interleaved French+English vs. French-only, and (3) Rust vs. Rust+English, we demonstrate that mixing English with structurally rich languages degrades both perplexity and structural pattern acquisition. At 125M parameters, French achieves 100% grammar probe accuracy in∼197M tokens while English remains at chance after 3B tokens—a >15x efficiency gap Wasserman 2026. The pollution effect is demonstrated directly by interleaved training: when French and English alternate in the same model, French grammar accuracy degrades from 100% to 50–70%—English does not merely fail to help, it actively corrupts the French grammatical signal. The same pattern appears in code: when Rust is interleaved with English text, the model shows 10x worse perplexity and 20% lower probe accuracy compared to Rust-only training. These findings suggest that English’s morphological poverty creates noise that pollutes learning of any structurally explicit system, whether natural language (French) or synthetic (Rust). Critically, we show that perplexity and probe accuracy are orthogonal dimensions: models can achieve identical accuracy with 11x different perplexity, revealing that standard metrics miss the depth of structural understanding. We propose that the field’s reliance on English-dominated training corpora may be a hidden bottleneck in language model capability development.
Adam Zachary Wasserman (Mon,) studied this question.