The first stage shifts the model's reasoning language from English to Italian through supervised fine-tuning on reasoning traces from toroe/ReasonXL-SFT.

The corresponding SFT model is:

DGurgurov/SmolLM3-3B-SFT-IT

Stage 2 — RL

This model applies RL (Dr. GRPO) to the corresponding SFT model.

The objective is to recover reasoning quality lost during supervised fine-tuning while preserving compliance with the target reasoning language.

Training uses a composite reward over verifiable mathematical problems.

Model Details

Base model: HuggingFaceTB/SmolLM3-3B
Target reasoning language: Italian
SFT dataset: toroe/ReasonXL-SFT
Training pipeline: SFT → Dr. GRPO
SFT model: DGurgurov/SmolLM3-3B-SFT-IT
This model: GRPO

Full training details, reward formulation, evaluation results, and methodology will follow soon.

Citation

If you use this model, please cite:

bibtex

@misc{gurgurov2026reasonxlshiftingllmreasoning,
      title={ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance},
      author={Daniil Gurgurov and Tom Röhr and Sebastian von Rohrscheidt and Josef van Genabith and Alexander Löser and Simon Ostermann},
      year={2026},
      eprint={2604.12378},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.12378},
}

DGurgurov/SmolLM3-3B-SFT-GRPO-IT

SmolLM3-3B-SFT-GRPO-IT

Stage 1 — SFT

Stage 2 — RL

Model Details

Citation

DGurgurov/
SmolLM3-3B-SFT-GRPO-IT