Skip to main content
Back

ReasonXL Teaches AI to Think in French or German Without Losing Accuracy

A new parallel corpus of 2M+ reasoning traces per language and a two-stage recipe let a 3B model reason natively in French, German, Italian, or Spanish without losing accuracy.

  • ReasonXL releases a parallel reasoning corpus of 2M+ samples per language across English, German, French, Italian, Spanish.
  • Data and models are available on HuggingFace under an actively expanding release.
  • Two-stage recipe (SFT then Dr. GRPO) shifts SmolLM3-3B's reasoning language while matching or beating baseline accuracy.
  • SFT alone drops science scores 5 to 10 points; RL with a composite reward fully recovers and often exceeds them.
  • Activation patching shows layers 6 to 8 form a causal bottleneck that decides output language identity.
  • English general knowledge drops by only 0.6 to 3.2 points after full adaptation, cross-lingual transfer largely preserved.

Most reasoning-tuned language models have a quiet habit: even when you ask a question in French or German, they slip into English inside their <think> block before answering. A new project called ReasonXL attacks that behavior with a large parallel corpus and a two-stage post-training recipe that reroutes a model's inner monologue into the language of the user, without giving up accuracy.

The English monologue problem

Recent multilingual reasoning models default to English for intermediate steps even when prompted in other languages, and they consistently score higher in English than elsewhere. That mismatch matters for anyone deploying LLMs outside anglophone markets: users cannot inspect a chain of thought written in a language they do not speak, and routing through English introduces translation-shaped errors in the semantics of the problem itself.

The authors argue that two ingredients have been missing: large-scale non-English reasoning data, and evidence that language-specific post-training can match English-level performance. Prior approaches either pivot through English, prompt-engineer the reasoning language (trading accuracy for compliance), or fine-tune on small in-language datasets that leave a visible gap.

A 44-billion-token parallel corpus

ReasonXL is the concrete data contribution: a parallel corpus of reasoning traces spanning English, German, French, Italian, and Spanish, with over two million aligned samples per language, each containing prompt, reasoning trace, and final output. The current release comes to roughly 9B tokens per language (44B total), and the authors say the translation pipeline is being extended toward 20 to 30B tokens per language.

The English source is stitched together from ten existing reasoning datasets, including Olmo, NVIDIA, Bercovich et al., and Wang et al., then filtered with Propella-1-4B, a small annotator model that scores 18 properties per document. The pipeline enforces:

  • Safety and integrity constraints (no PII, complete content, presence of reasoning indicators).