Language as a Latent Variable for Reasoning Optimization

The deeper claim here isn't that some languages produce better answers; it's that language selection functions as an implicit architectural choice during inference, meaning the same model weights can behave like meaningfully different reasoners depending on which language token stream they're operating in.

This connects directly to two threads in recent coverage. The cultural bias paper ('Why are all LLMs Obsessed with Japanese Culture') already showed that language choice and training data composition shape which topics models prioritize — this paper extends that observation from content bias into reasoning quality itself. More structurally, the DiffMAS work ('Learning to Communicate') argues that communication protocols between agents should be treated as learnable variables rather than fixed text. The polyGRPO finding is essentially the same argument applied to a single model's internal monologue: if you stop constraining the language of reasoning, the model finds better paths. The two papers arrive at a similar principle from opposite directions. The Global South multilinguality survey is also relevant context, since it documents how underrepresented languages remain undertrained — which would cap the gains polyGRPO can claim for those languages specifically.

If polyGRPO's gains hold when tested on languages with sparse training representation (below roughly 0.1% of pretraining data), the mechanism is genuinely architectural; if gains collapse there, the effect is better explained by training data density than by anything intrinsic to language structure.