Efficient Reasoning Speaks English: Reasoning Compression Anglicizes Weak-Language Thought

Reasoning-compression recipes report their token savings as a single number, measured in English, yet reasoning models are deployed multilingually. We evaluate two recipes—VeriThinker, an SFT recipe at 7B, and L1, a length-controlled RL recipe at 1.5B—across the 11 languages of MGSM. The reported savings prove to be an average over starkly unequal parts: compressing R1-Distill-7B saves 7% of thinking tokens in German and 15% in English, but 56–59% in Telugu and Thai. Accompanying this uneven saving is CoT anglicization. Compression lifts the English share of the chain of thought, raising the Latin-script share of Thai, Telugu, and Bengali reasoning from under 20% to 70–95%. It also anglicizes final answers, while leaving Chinese untouched. Evaluating the intermediate checkpoint of L1’s pipeline attributes both effects to the length-control step. The preceding English-math-RL stage instead lengthens reasoning and shifts the language of thought three times less. Finally, using L1’s prompt-controlled budgets as a within-checkpoint probe, we find budget elasticity concentrated in high-resource languages: raising the budget from 512 to 3,600 tokens buys 10–14 points in English, Chinese, and French, and nothing in Thai, Telugu, Bengali, or Swahili. Accuracy never falls, but both the savings and the shift away from native-language reasoning land hardest on the weakest languages—a distributional effect that a single English number hides.