PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca, Daniel D'souza, Alexandre Berard, Thomas Euyang, Marzieh Fadaee, Julia Kreutzer

arXiv:2609.10445v1cs.CL

TL;DR

Multilingual reasoning remains largely English-centric, risking lost intent and missing target-language knowledge. The paper studies data composition and scheduling for multilingual SFT, finding that broader language coverage, multilingual non-reasoning data, and sufficient English reasoning supervision support transfer across languages.

  • Problem

    Reasoning models often reason in English despite non-English prompts, creating a gap between multilingual understanding and in-language reasoning and risking loss of source-language intent or knowledge.

  • Method

    The paper builds a dual-mode multilingual reasoning model by combining English reasoning data, multilingual non-reasoning data, and translated target-language reasoning traces through data mixing and scheduling.

  • Results

    Broader language coverage improves transfer without interference, while multilingual instruction data benefits L2 reasoning and reasoning performance, including on languages without direct reasoning supervision.

  • Takeaways & Limitations

    Reasoning capability and reasoning language can be learned as separable properties, making reasoning language transferable rather than requiring language-by-language construction.

  • Takeaways & Limitations

    Translated reasoning supervision anchors reasoning style, cultural framing, and problem decomposition to English templates, with lower translation quality for low-resource languages.

Abstract

from arXiv · show

Reasoning language models have made substantial advances on a variety of complex tasks, yet their capabilities remain overwhelmingly English-centric: models primarily reason in English regardless of the language they are prompted in. This is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge more readily expressed in the target language. In this work, we advance L2 reasoning, the ability of a model to reason consistently in the language of the user's prompt, thus building an in-language bridge between the prompt and the answer. We approach this problem from a data-centric angle, investigating how to optimize data composition and scheduling in SFT for reasoning generalization. Building Tiny Aya L2-Thinker at 3.35B scale, we achieve an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. We show the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. These findings indicate that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing and without requiring reasoning supervision in every target language. We release our model weights and multilingual reasoning data to support further research on accessible, in-language reasoning.

1 Introduction

Reasoning models remain predominantly English-centric, risking loss of language-specific intent and knowledge. The paper investigates data composition and scheduling for transferring reasoning into users’ languages while preserving task performance.

  • Motivation: Reasoning models commonly answer non-English prompts with English reasoning, despite responding to prompts in the target language.
  • Motivation: English reasoning can lose source-language intent, nuance, framing, and culturally embedded knowledge, potentially lowering accuracy.
  • Motivation: Smaller models struggle to maintain both task performance and L2 reasoning, while larger L2 models still fall short of perfect in-language reasoning.
  • Contribution: Tiny Aya L2-Thinker extends a 3.35B multilingual base to dual-mode L2 reasoning across 45 languages.
  • Findings: The study shows that broader language coverage increases transfer without interference, including to languages lacking direct L2 reasoning supervision.
  • Findings: Multilingual non-reasoning data supports cross-lingual transfer, English reasoning supplies a backbone for difficult math, and data mixing improves accuracy–L2 reasoning trade-offs.

2 Methodology

The methodology treats multilingual reasoning as a data-composition problem. It combines translated reasoning traces, English reasoning, and multilingual instruction data during supervised fine-tuning to promote transfer beyond directly supervised languages.

  • Approach: The study focuses on SFT, using reference reasoning traces to build L2 reasoning without large datasets for every language.
  • Data augmentation: Translate-train converts abundant English reasoning traces into target-language supervision while preserving the original problem and solution structure.
  • Data augmentation: Translated reasoning is treated as scarce because long traces are costly to translate and errors can accumulate, especially in math, code, and lower-resource languages.
  • Generalization: The study asks whether mixed multilingual supervision can condition reasoning language on the input language, including languages without L2 reasoning traces.
  • Data mixing: Training combines English reasoning for task-solving capability, multilingual reasoning for target-language supervision, and cheaper multilingual non-reasoning instruction data for language alignment.

3 Experimental Setup

Experiments use the 3.35B multilingual Tiny Aya family, multilingual reasoning and non-reasoning mixtures, and evaluations spanning reasoning, instruction following, generation, and cultural capabilities across supported languages.

  • Base model: Tiny Aya is a 3.35B multilingual family supporting more than 70 languages across five world regions.
  • Base model: The model is extended to 32K-token context through continued training with interleaved 8K and 32K sequences.
  • Training data: English reasoning data covers mathematics, science, and general reasoning, while translated multilingual reasoning data supervises 45 languages including English.
  • Training data: Multilingual non-reasoning data spans approximately 4.9M samples across 67 languages, using empty thinking blocks to preserve ordinary instruction following.
  • Evaluation: Evaluation covers math, multicultural reasoning, localized instruction following, and multilingual open-ended generation across 60 supported languages.
  • Metrics: The primary metrics are task accuracy and L2 reasoning rate, which measures whether reasoning is predominantly written in the prompt language.

4 Results

The results compare trained in-language reasoning with language forcing across task performance, reasoning efficiency, and language-resourcedness. Tiny Aya L2-Thinker provides broad, stable L2 reasoning with concise traces and limited accuracy costs, though competition-level math remains a weakness.

  • Comparative setup: Tiny Aya L2-Thinker is evaluated against English-reasoning and L2 systems using English and other-language metrics averaged across benchmark languages.The comparison includes language forcing for Qwen3.5-4B and trained L2 systems including Tiny Aya L2-Thinker, M-Thinker-7B, and Magistral-Small-24B.
  • L2 reasoning performance: Above 93% of Tiny Aya L2-Thinker’s traces use the target language across every benchmark, including seen and unseen languages, with consistently low variance.It exceeds M-Thinker-7B on five of six benchmarks at half the size and exceeds Magistral-Small-24B on all tasks.
  • Accuracy trade-offs: Switching from English to in-language reasoning reduces accuracy by at most 2–3% on five tasks, while PolyMath falls from 18.6% to 11.1%.English performance is preserved on most benchmarks, with the clearest drop being 5 points on GlobalPIQA.
  • Accuracy trade-offs: Qwen3.5-4B leads PolyMath at 40.3%, but the gap to Tiny Aya L2-Thinker partly reflects the absence of reinforcement-learning refinement in the authors’ models.Math-targeted reinforcement learning improves M-Thinker-7B’s PolyMath score but coincides with much weaker open-ended generation.
  • Efficiency: Longer reasoning traces are associated with higher 4-gram repetition, while Tiny Aya L2-Thinker generally uses fewer tokens and shows less repetition than Qwen3.5-4B.PolyMath is the exception where harder competition-level problems warrant more deliberation.
  • Low-resource coverage: 94.3% and 94.5% are Tiny Aya L2-Thinker’s L2 reasoning rates on tiers 3 and 4, while M-Thinker-7B falls from 99.6% to 71.6%.Tiny Aya L2-Thinker remains strong across L2 coverage, task accuracy, and reasoning-trace efficiency as language resourcedness declines.
  • Overall comparison: Tiny Aya L2-Thinker covers 45 languages versus M-Thinker-7B’s 6, with roughly half its L2-rate variance and more efficient traces than Qwen3.5-4B.The comparison describes Tiny Aya L2-Thinker as maintaining comparable performance while improving coverage and reasoning efficiency.

5 Analysis and Building Blocks

The analysis tests how language coverage, multilingual non-reasoning data, and English reasoning data affect L2 reasoning transfer and task performance. Broader coverage selectively improves unseen-language reasoning, while data mixing balances reasoning-language alignment with accuracy more reliably than sequential alternatives.

  • Language coverage: The coverage experiment compares one-language specialists, two-language regional models, and one jointly trained all-languages model with region-balanced evaluation.Seen and unseen language sets are fixed in advance for each benchmark.
  • Language coverage: Broader language coverage raises unseen-language L2 reasoning while leaving English, seen-language accuracy, and seen-language reasoning essentially stable.Across MGSM, Marco-Bench-MIF, and GlobalPIQA, unseen-language L2 reasoning rises 14 →35 →60, 10 →13 →16, and 19 →21 →42, respectively.
  • Non-reasoning data: A 10% multilingual non-reasoning fraction raises MGSM unseen-language L2 reasoning from 46% to 89% and task accuracy from 49% to 67%.These gains occur at matched training conditions while L2 reasoning remains fixed at 10% of the mixture.
  • Non-reasoning data: Non-reasoning data works best around 20–30%; beyond that, the model increasingly produces empty reasoning traces.The data couples input language to output language, and this anchoring transfers into reasoning mode.
  • English reasoning backbone: Math benchmarks benefit from a heavier English reasoning backbone, whereas open-ended tasks gain little from additional English data and show more complex L2 reasoning-rate changes.For math, accuracy climbs while L2 reasoning stays nearly flat; for open-ended tasks, 10% English behaves similarly to 100%.
  • Training strategies: Data mixing offers the best trade-off between task accuracy and L2 reasoning, while merging collapses toward English reasoning and sequential training sacrifices accuracy.Sequential training also introduces unexpected language confusion, whereas mixing provides a more predictable fallback to English reasoning when L2 reasoning fails.
  • Data pillars: English reasoning alone reaches 36.7% task accuracy and 12.8% L2 reasoning, multilingual reasoning raises L2 reasoning to 86.1%, and non-reasoning data improves both axes further.The three data pillars are English reasoning, multilingual reasoning, and multilingual non-reasoning data.

6 Related Work

Related work addresses multilingual reasoning through English-centered bridges, inference-time language control, translated training data, reinforcement learning, and model recomposition. This paper distinguishes its target—fully in-language reasoning—from methods that transfer accuracy, optimize hybrid traces, or rely on language-specific specialists.

  • English-centered approaches: English-centered bridge methods improve multilingual input handling or mathematical accuracy while leaving intermediate reasoning predominantly in English.Their transferred capability is reasoning accuracy rather than reasoning in the user’s language.
  • Mixed-language reasoning: Hybrid-trace methods retain an English scaffold or deliberately code-switch, whereas this work targets traces readable end to end in the prompt language.The cited hybrid approaches optimize mixed-language reasoning rather than monolingual target-language traces.
  • Inference-time control: Reasoning models show lower input–output language consistency than general-purpose models, with the reasoning process identified as the weakest point.Reported general-purpose consistency exceeds 95% across 18 languages, while reasoning models are substantially lower.
  • Inference-time control: Inference-time language forcing mainly redistributes probability toward target-language reasoning and succeeds only when the model can already sustain extended reasoning there.Models follow the directive more reliably in final answers than in the preceding reasoning trace.
  • Translated supervision: Translate-train work commonly translates short mathematical chains into a few languages, whereas this paper translates long-form traces across 44 languages and three domains as a scarce seed.The paper keeps fewer than 5K samples per language to study crosslingual generalization beyond covered languages.
  • Reinforcement learning: Reinforcement-learning studies directly optimize reasoning language or language consistency, but language-consistency rewards can impose measurable accuracy costs and effects vary across languages.One cited result reports severe, largely irreversible English drift for lower-resource languages as accuracy rises.
  • Training regimes: Joint multilingual reasoning training is contrasted with specialist regimes, where language-consistency enforcement can hurt crosslingual generalization but parallel multilingual scaling improves it.The comparison emphasizes that conclusions depend on the training regime.
  • Data resources: The paper releases multilingual reasoning data covering 44 languages besides English to address the scarcity of non-English reasoning traces.The cited dataset landscape remains dominated by English-only reasoning corpora, with only limited multilingual coverage.

7 Conclusion

The paper argues that multilingual reasoning can be optimized by separating reasoning capability from reasoning language. Large-scale English reasoning supervision, combined with multilingual supervision and broader instruction data, supports transferable in-language reasoning despite limited target-language reasoning traces.

  • Multilingual reasoning can be achieved by pairing a small fraction of non-reasoning data with a sufficiently large English reasoning dataset.
  • Reasoning capability and reasoning language are related but separable: complex-problem solving can be learned from reasoning supervision, while multilingual supervision teaches its language of expression.
  • Reasoning language can generalize to languages without direct reasoning supervision through limited multilingual reasoning data and broader multilingual instruction.
  • The proposed separation offers a scalable path toward models that reason in the language of their users across languages.

8 Limitations

The paper identifies three limitations: translated reasoning traces may inherit English-centric artifacts, automatic metrics do not assess reasoning quality, and reported performance depends on an explicit prompting trigger.

  • Translated reasoning traces anchor reasoning style, cultural framing, and problem decomposition to English templates.For low-resource languages, translation quality may degrade and may not reflect natural approaches to multi-step deduction.
  • Automatic evaluation measures linguistic compliance rather than reasoning depth, coherence, usefulness, fluency, or cultural appropriateness.The study does not conduct human evaluations of reasoning-trace quality.
  • Reported L2 reasoning rates rely on an explicit instruction to think in the prompt's language.Without this trigger, the model may revert to English, and constrained or mixed-language deployments may see degraded performance.

A Training

The training pipeline extends Tiny Aya Base to longer contexts and fine-tunes it with a filtered, packed multilingual SFT mixture using standard next-token prediction.

  • Tiny Aya Base's context length is extended from 8K to 32K tokens in a single training stage.
  • The context-extension stage uses 12,000 steps on an interleaved 8K/32K mixture with a 3:1 ratio and length buckets spanning 8K to 32K.A linear learning-rate schedule begins at 1.25 × 10^-4.
  • Tiny Aya L2-Thinker is fine-tuned with a standard next-token crossentropy objective for four epochs over the packed SFT mixture.Training uses 32 NVIDIA H100 GPUs, a global batch size of 32, Adam optimization, weight decay, gradient clipping, cosine decay, and warmup.
  • The English reasoning dataset is filtered to remove translation-prone examples, including code-switching, translation discussion, target-language mentions, and fragile formatting constraints.Language identification uses FastText with GlotLID as a fallback.

B.2 Translated Reasoning Data

The translated reasoning training data spans 44 languages besides English, multiple regions, and three broad domains, with different sampling caps for controlled experiments and the final model.

  • The training data contains translated reasoning examples across 44 languages besides English, covering Europe, Asia-Pacific, West Asia, Africa, and South Asia.
  • The data spans Math, Science, and General domains.
  • Controlled experiments cap samples at 5K per language, whereas the final Tiny Aya L2-Thinker model uses all available data.

C Benchmarking Details

The evaluation uses multilingual benchmarks with balanced seen-versus-unseen language splits, while restricting languages to those supported and identifiable for Tiny Aya. Each example receives one completion under a common 32K-token context, with reasoning separated from the final answer.

  • Language splits: Seen languages occur in the ten-language training set, whereas unseen languages are held out and selected with balanced regional coverage.Where no natural unseen language exists, Urdu substitutes for West Asia or South Asia in specified benchmarks.
  • Language coverage: The benchmark language lists are restricted to languages supported by Tiny Aya and identifiable by FastText or GlotLID.Qwen3.5-4B supports more languages, but its exact language coverage is undisclosed.
  • Language coverage: PolyMath, Marco-Bench-MIF, MIST-OEG, and Macaron-MCQ cover more than 90% of their original benchmark languages, while MGSM and GlobalPIQA use diverse subsets.The subsets still include both high-resource and low-resource languages.
  • Evaluation protocol: The study generates one completion per example, separates reasoning from the final answer using model-specific conventions, and evaluates all models with a 32K-token context window.The conventions include distinct thinking delimiters for Tiny Aya, Qwen, M-Thinker, and Magistral.

E Benchmark results by language

The analysis measures per-language accuracy and L2 reasoning rates across benchmarks and examines whether longer reasoning traces coincide with redundant thinking. Tiny Aya L2-Thinker is described as more efficient than Qwen3.5-4B under language forcing, while allocating extra tokens to hard math.

  • Evaluation metrics: Per-language benchmark results report task accuracy and L2 reasoning rate, with means and standard deviations over languages.PolyMath additionally separates medium, high, and top difficulty levels and uses a weighted average in the main section.
  • Thinking efficiency: Qwen3.5-4B with thinking-prefix language forcing produces systematically longer traces, with many samples exceeding 100 tokens and accumulating at the maximum budget.Longer traces are tightly linked to higher doomlooping scores for Qwen and Tiny Aya, whereas Magistral’s relationship is weaker.
  • Thinking efficiency: Tiny Aya L2-Thinker generates more efficient thinking traces, spending extra tokens mainly on the hard PolyMath set.The comparison uses per-sample traces across languages and relates reasoning length to 4-gram repetition.

G Ablations

The ablation compares joint mixing, sequential adaptation, and linear merging under a controlled multilingual setup. Joint mixing is never dominated: it preserves a strong accuracy–L2 reasoning balance and a predictable English fallback on unseen languages.

  • Ablation design: The comparison reuses a ten-language, five-region setup to test joint mixing against sequential adaptation and separately trained specialist merging.All three strategies combine English reasoning with L2 supervision under a fixed training budget.
  • Accuracy–L2 trade-off: Merging has higher task accuracy but collapses in-language reasoning, with seen-language L2 rates of 59/17/49 versus 99/95/100 for mixing.The reported values are ordered MGSM, Marco-Bench-MIF, and GlobalPIQA.
  • Accuracy–L2 trade-off: Sequential adaptation reaches the prompt language more often on held-out inputs but lowers accuracy by forgetting part of the base English capability.Mixing gives up some unseen-language L2 reasoning for higher accuracy than sequential adaptation and is best on both axes for unseen MGSM.
  • Fallback behavior: Joint mixing keeps English as a stable fallback on unseen languages, whereas sequential adaptation scatters fallback across other languages and merged models reason primarily in English.Figure 12 organizes thinking-language distributions by region and benchmark across four models.
  • Accuracy–L2 trade-off: Mixing is never dominated across the task accuracy–L2 reasoning rate trade-off.Figure 11 compares the strategies on seen and unseen languages across MGSM, Marco-Bench-MIF, and GlobalPIQA.
  • Evaluation scope: The evaluation suite spans multilingual benchmarks including Macaron-MCQ, PolyMath, and multiple task-specific datasets, with region-balanced seen and unseen splits.The tables report per-language accuracy and L2 reasoning rates for benchmark-specific language sets.

H Reasoning Errors

The error analysis contrasts Tiny Aya L2-Thinker with Qwen3.5-4B and Magistral-Small-24B on multilingual prompts. Tiny Aya maintains language consistency, while the other models exhibit doomlooping, code-switching, or mixed-language generation in the examples shown.

  • Language consistency: Tiny Aya L2-Thinker reasons and generates in Korean with perfect language consistency in the MIST-OEG example, while Magistral mixes German, English, and Korean.Tiny Aya also achieves the higher overall task score in that example.
  • Code-switching: Magistral-Small-24B frequently code-switches, while Tiny Aya L2-Thinker does not.The example reasoning and generation include mixed-language segments in Magistral’s output.
  • Code-switching: In the wedding-invitation example, Magistral’s reasoning proposes a vague non-financial explanation and its generation contains German and English fragments alongside Korean.The reasoning explicitly plans a polite refusal without directly mentioning travel costs.
  • Doomlooping: Qwen3.5-4B reasons primarily in English, enters a doomloop, and produces no answer on the Korean example, whereas Tiny Aya completes it consistently.The Qwen trace repeatedly revisits whether to answer in Korean before failing to generate.
Loading 2609.10445v1…