ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs

Daniil Gurgurov; Tom Röhr; Sebastian von Rohrscheidt; Josef van Genabith; Alexander Löser; Simon Ostermann

cs.CLApr 14, 2026

ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs

Authors: Daniil Gurgurov, Tom Röhr, Sebastian von Rohrscheidt, Josef van Genabith, Alexander Löser, Simon Ostermann

Organizations: German Research Center for Artificial Intelligence (DFKI), Saarland University · Berliner Hochschule für Technik

Abstract

Despite advances in multilingual capabilities, most large language models (LLMs) remain English-centric in their training and, crucially, in the reasoning traces they produce: even for non-English problems, they predominantly reason in English, leaving non-English users with intermediate reasoning they cannot follow. We address this disparity with two contributions. (a) We introduce $\texttt{ReasonXL}$ , the first large-scale parallel corpus of cross-domain reasoning traces in five European languages (English, German, French, Italian, and Spanish): over 2.5M aligned samples (11B tokens) per language, each comprising prompt, reasoning trace, and final output, enabling direct supervision of language-specific reasoning. (b) Using $\texttt{ReasonXL}$ , we demonstrate that LLMs can be adapted to reason entirely in a target language through a simple two-stage pipeline of supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR), matching or exceeding the base models' English-reasoning performance in two of four languages and recovering most of it in the remaining two.

Figures & tables

Panel of Table 1 — Table 1: Dataset construction pipeline (far left), per-language corpus size and sequence lengths (top-middle), composition (bottom-middle), and property distributions (top-right). Dataset sources: 1 Wang et al. (2025a) , 2 Olmo et al. (2025) , 3 Bercovich et al. (2025) , 4 NVIDIA (2025) . Bottom-right: English performance of SmolLM3-3B pre- and post-SFT: consistent gains demonstrate the dataset’s potential to push strong baselines further.

		SmolLM3-3B				OpenThinker3-7B
Lang	Method	Sci.	Lang.	Ov.	%TL	Sci.	Lang.	Ov.	%TL
de	Base	$61.6$	$65.5$	$63.5$	$0.0$	$\boldsymbol{61.3}$	$58.4$	$59.9$	$0.0$
	+ SFT	$50.9$	$60.7$	$55.8$	$96.5$	$33.7$	$51.6$	$42.6$	$87.3$
	+ GRPO	$\boldsymbol{61.6}$	$\boldsymbol{65.9}$	$\boldsymbol{63.7}$	$100.0$	$57.7$	$\boldsymbol{63.8}$	$\boldsymbol{60.7}$	$100.0$
it	Base	$62.2$	$67.6$	$64.9$	$0.0$	$\boldsymbol{61.3}$	$65.4$	$63.3$	$0.0$
	+ SFT	$58.1$	$71.2$	$64.6$	$96.2$	$35.8$	$52.6$	$44.2$	$92.0$

Table 2: Domain-averaged performance on target-language tasks. SFT steers the model to reason in the target language; GRPO recovers performance while preserving compliance. Sci. : MATH100/GPQA/MGSM; Lang. : Belebele/GMMLU/INCLUDE; Ov. : all six tasks, each averaged over 5 runs.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 1 — Figure 1: Translation prompt template. {field} is one of: input, thinking/reasoning, or output.

Lang	Sample
en	Input: Which of the following statements is true about the singularity at the center of a black hole? Reasoning: Okay, I need to figure out which of these options is correct about the singularity at the center of a black hole. Output: The answer is (C). Explanation: A singularity at the center of a black hole is a region where the spacetime curvature becomes infinite..
de	Input: Welcher der folgenden Aussagen ist in Bezug auf die Singularität im Zentrum eines Schwarzen Lochs wahr? Reasoning: Okay, ich muss herausfinden, welche dieser Optionen in Bezug auf die Singularität im Zentrum eines Schwarzen Lochs korrekt ist. Output: Die Antwort ist (C). Erklärung: Eine Singularität im Zentrum eines Schwarzen Lochs ist eine Region..
es	Input: ¿Cuál de las siguientes afirmaciones es verdadera sobre la singularidad en el centro de un agujero negro? Reasoning: Bueno, necesito averiguar cuál de estas opciones es correcta acerca de la singularidad en el centro de un agujero negro. Output: La respuesta es (C). Explicación: Una singularidad en el centro de un agujero negro es una región donde la curvatura del espacio-tiempo se vuelve infinita..
fr	Input: Laquelle des affirmations suivantes est vraie concernant la singularité au centre d’un trou noir ? Reasoning: D’accord, je dois déterminer laquelle de ces options est correcte concernant la singularité au centre d’un trou noir. Output: La réponse est (C). Explication : Une singularité au centre d’un trou noir est une région où la courbure de l’espace-temps devient infinie..
it	Input: Quali delle seguenti affermazioni è vera riguardo alla singolarità al centro di un buco nero? Reasoning: Okay, devo capire quale di queste opzioni è corretta riguardo alla singolarità al centro di un buco nero. Output: La risposta è (C).Spiegazione: Una singolarità al centro di un buco nero è una regione in cui la curvatura dello spaziotempo diventa infinita…

Appendix

Table 3: Example translations from the parallel Reason XL corpus.

Technical Content Class	Sampling Ratio
code_heavy_math_heavy	0.60
math_heavy	0.30
non_technical	0.50
basic_technical	0.80

Appendix

Table 4: Class-aware sampling ratios used during dataset balancing.

Category	Configuration
Decoding Strategy	Nucleus Sampling
Temperature	0.1
Top- $p$	1.0
Top- $k$	$-1$
Max Length	6,144 tokens

Appendix

Table 5: Decoding and generation parameters used for translation.

Measurement	EN source	Translations
Samples passing all checks	97.2%	76.6–77.2%
Samples repaired	0.4%	17.2%
Leaked instruction block	0%	3.1–4.4%
Upper-cased prose	0.005%	0.22–5.42%

Appendix

Table 6: Rates on the untranslated English control vs. the translated splits under identical checks.

Property	Aggregated category	%
Content quality	good or excellent	98.0
	poor or unacceptable	0.6
Information density	dense or adequate	94.0
Educational value	high or moderate	88.0
Reasoning orientation	analytical or explanatory	96.0
	basic or minimal	$<4$

Appendix

Table 7: Property distribution of the re-annotated German split (

{\sim}2.5

M samples). Percentages are over annotated documents in the split.

Hyperparameter	Value
Epochs	2
Max sequence length	16,384
Packing	Enabled
Precision	bfloat16
Optimizer	adamw_torch_fused
Per-device batch size	4

Appendix

Table 8: SFT training hyperparameters (shared across all languages).

Hyperparameter	Value
Loss type	Dr. GRPO
Epochs	1
Generations per prompt ( $G$ )	8
Max completion length	8,192
Sampling temperature	1.0
Top- $p$	1.0

Appendix

Table 9: Dr. GRPO training hyperparameters.

Reward Component	DE	IT	ES	FR
Accuracy	1.0	1.0	1.0	1.0
Language	0.2	0.2	0.2	0.2
Format	0.1	0.1	0.1	0.1
Repetition	0.3	0.3	0.3	0.3

Appendix

Table 10: Reward weights per language.

Source datasets
Artifact	License
Nemotron-Cascade-SFT	CC-BY-4.0
Dolci-Think-SFT (Olmo 3)	ODC-BY
Llama-Nemotron-PTD	CC-BY-4.0
Nemotron-Science-v1	CC-BY-4.0
Nemotron-IF-Chat-v1	CC-BY-4.0

Appendix

Table 11: Licenses of artifacts used in this work.

		Science			Lang.
Lang	Method	MATH100	GPQA	MGSM	Bele.	GMMLU	INCL.	Avg.	%TL
de	SmolLM3	$66.1{\scriptstyle\pm 1.4}$	$\boldsymbol{36.9{\scriptstyle\pm 3.5}}$	$\boldsymbol{81.8{\scriptstyle\pm 2.1}}$	$81.5{\scriptstyle\pm 1.3}$	$\boldsymbol{66.5{\scriptstyle\pm 0.7}}$	$\boldsymbol{48.5{\scriptstyle\pm 1.3}}$	$63.5$	$0.0$
	+ SFT	$51.5{\scriptstyle\pm 2.1}$	$30.5{\scriptstyle\pm 3.0}$	$70.8{\scriptstyle\pm 2.3}$	$82.7{\scriptstyle\pm 0.5}$	$55.2{\scriptstyle\pm 1.9}$	$44.3{\scriptstyle\pm 1.3}$	$55.8$	$96.5$
	+ GRPO	$\boldsymbol{70.5{\scriptstyle\pm 2.7}}$	$34.0{\scriptstyle\pm 2.8}$	$80.4{\scriptstyle\pm 1.2}$	$\boldsymbol{85.3{\scriptstyle\pm 0.7}}$	$66.4{\scriptstyle\pm 2.0}$	$45.8{\scriptstyle\pm 2.6}$	$\boldsymbol{63.7}$	$100.0$
it	SmolLM3	$64.4{\scriptstyle\pm 0.8}$	$\boldsymbol{39.4{\scriptstyle\pm 3.2}}$	$82.9{\scriptstyle\pm 1.7}$	$75.9{\scriptstyle\pm 0.7}$	$63.7{\scriptstyle\pm 1.6}$	$63.1{\scriptstyle\pm 1.2}$	$64.9$	$0.0$
	+ SFT	$58.6{\scriptstyle\pm 1.8}$	$33.8{\scriptstyle\pm 1.2}$	$82.0{\scriptstyle\pm 1.4}$	$84.0{\scriptstyle\pm 1.3}$	$65.5{\scriptstyle\pm 0.8}$	$64.1{\scriptstyle\pm 3.2}$	$64.6$	$96.2$

Appendix

Table 12: Performance on downstream target-language tasks for SmolLM3-3B .

		Science			Lang.
Lang	Method	MATH100	GPQA	MGSM	Bele.	GMMLU	INCL.	Avg.	%TL
de	OpenThinker3-7B	$\boldsymbol{69.7{\scriptstyle\pm 1.3}}$	$\boldsymbol{44.3{\scriptstyle\pm 1.0}}$	$70.0{\scriptstyle\pm 1.2}$	$78.1{\scriptstyle\pm 1.0}$	$59.3{\scriptstyle\pm 0.9}$	$37.8{\scriptstyle\pm 3.1}$	$59.9$	$0$
	+ SFT	$25.1{\scriptstyle\pm 2.8}$	$25.4{\scriptstyle\pm 4.1}$	$50.6{\scriptstyle\pm 2.6}$	$62.5{\scriptstyle\pm 1.4}$	$47.4{\scriptstyle\pm 1.6}$	$44.9{\scriptstyle\pm 2.4}$	$42.6$	$87.3$
	+ GRPO	$63.0{\scriptstyle\pm 2.0}$	$28.6{\scriptstyle\pm 1.5}$	$\boldsymbol{81.4{\scriptstyle\pm 0.8}}$	$\boldsymbol{78.6{\scriptstyle\pm 0.9}}$	$\boldsymbol{61.0{\scriptstyle\pm 0.9}}$	$\boldsymbol{51.7{\scriptstyle\pm 1.7}}$	$\boldsymbol{60.7}$	$100.0$
it	OpenThinker3-7B	$\boldsymbol{67.5{\scriptstyle\pm 1.6}}$	$\boldsymbol{42.2{\scriptstyle\pm 2.1}}$	$74.2{\scriptstyle\pm 0.6}$	$79.0{\scriptstyle\pm 1.1}$	$59.3{\scriptstyle\pm 2.5}$	$57.8{\scriptstyle\pm 2.6}$	$63.3$	$0$
	+ SFT	$27.5{\scriptstyle\pm 4.0}$	$24.7{\scriptstyle\pm 1.7}$	$55.0{\scriptstyle\pm 2.3}$	$65.8{\scriptstyle\pm 1.5}$	$47.0{\scriptstyle\pm 2.5}$	$45.2{\scriptstyle\pm 2.6}$	$44.2$	$92.0$

Appendix

Table 13: Performance on downstream target-language tasks for OpenThinker3-7B .

Explore similar work

x1: Learning to Think Adaptively Across Languages and Cultures
Apr 18, 2026Yangfan Ye, Xiaocheng Feng, Xiachong Feng +8Multilingual Language Models Cross-Lingual Reasoning in Language Models
Harbin Institute of Technology · Peng Cheng Laboratory · The University of Hong Kong +1
Rethinking the Multilingual Reasoning Gap with Layer Swap
May 26, 2026Maxence Lasbordes, Amélie Chatelain, Djamé SeddahMultilingual Language Models Multilingual Language Model Evaluation
1LightOn, Paris · 2Inria, Paris
Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning
Sep 9, 2026Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca +5Multilingual Language Models Cross-Lingual Reasoning in Language Models
Cohere Labs · Cohere

Apr 18, 2026J/K move · Enter open · S save

x1: Learning to Think Adaptively Across Languages and Cultures

Yangfan Ye, Xiaocheng Feng, Xiachong Feng, Yichong Huang+7

Harbin Institute of Technology · Peng Cheng Laboratory · The University of Hong Kong+1

Multilingual Language Models Cross-Lingual Reasoning in Language Models Cross-Lingual Transfer Large Reasoning Models

Open paper AskPDF arXiv

Languages encode distinct abstractions and inductive priors, yet most large language models (LLMs) overlook this diversity by reasoning in a single dominant language. In this work, we introduce x1, a family of reasoning models that can adaptively reason in an advantageous language on a per-instance basis. To isolate the effect of reasoning-language choice, x1 is constructed without expanding the model's knowledge boundaries and is trained by contrasting linguistically distinct reasoning trajectories for the same input. Our extensive experiments demonstrate the benefits of adaptive multilingual reasoning across multilingual mathematical reasoning and culturally grounded tasks. Moreover, our results challenge a simplistic view of scaling laws: while scaling reduces cross-lingual disparities in procedural domains such as math reasoning, it does not eliminate the advantages of culture-associated languages in culturally grounded tasks, as we empirically show that such reasoning enables more efficient and accurate cultural knowledge recall. Overall, our findings establish language choice as a functional component of reasoning, with implications for building more generalist and globally competent reasoning models.

Abstract

Explore similar work

x1: Learning to Think Adaptively Across Languages and Cultures

Rethinking the Multilingual Reasoning Gap with Layer Swap

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning