Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation

Investigating the Role of Reasoning-Language Alignment in Monolingual Retrieval-Augmented Generation. By Oliver Hauck and two others, from Johannes Gutenberg University Mainz and the University of Colorado Boulder. Published on the second of October. Its code repository is worked out: about fifty source files, with a readme.

Large language models reason better when they produce reasoning traces, but current models are trained to reason mostly in English. Earlier work showed that forcing a model to reason in another language lowers accuracy, even when that language matches the language of the prompt. That result came from a setting with a short prompt. This study asks whether it still holds for retrieval-augmented generation, where the model has to read and integrate a large amount of retrieved evidence written in the target language. The answer it reports is a qualified no. When the query and the documents are in German, making the model reason in German helps, and the reason is alignment between the languages, not the model's skill in any one of them.

To test this, the authors built a fully monolingual German question-answering testbed set in the fictional world of The Dark Eye, a tabletop role-playing game. The choice is deliberate. The setting is richly documented in German, but niche enough that a model cannot answer from memory, so it has to rely on what it retrieves.

On this testbed they vary the language an agentic retrieval system is forced to reason in. Aligning the reasoning language with the language of the query and the retrieved documents helps. Forced German reasoning beats forced French reasoning, even though the model scores higher on benchmarks in French. That contrast is the heart of the argument: if proficiency drove the result, French should have won, so the benefit has to come from alignment. The advantage also grows when the retrieved context is richer and structure-aware, which is what one would expect if the gain comes from working closely with the evidence in its own language.

The finding has a ceiling. Forced German reasoning only reaches the level of the model's native, unconstrained English reasoning. It does not surpass it. So alignment recovers ground, but it does not open new ground, and the authors read this as showing that native multilingual reasoning is needed rather than a forced switch at inference time. The evidence also comes from one language, German, one niche domain, and one agentic retrieval setup, with French as the comparison language. Those are the conditions under which the claim was measured.

The result reframes an earlier negative finding. Forcing a non-English reasoning language looked like a pure cost when the prompt was short. Once the model has to work through a large body of evidence in that language, matching it becomes an advantage, though one that so far only catches up with English. The testbed and the question-answering benchmark are released publicly.