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

AI summary

Overview

Research area: Multilingual natural language processing, reasoning language models, supervised fine-tuning data composition.

Technical level: Intermediate — requires familiarity with supervised fine-tuning (SFT), chain-of-thought reasoning traces, and standard multilingual NLP benchmarks, but the core ideas are explained accessibly.

Scope: A data-centric study showing how to compose and schedule SFT data so that a small (3.35B) multilingual model learns to reason in the user's prompt language across 45 languages without sacrificing task accuracy.

What This Paper Is About

Modern reasoning language models overwhelmingly "think" in English, regardless of the language the user writes in — a phenomenon the authors call the multilingual reasoning gap. This forces non-English speakers to trust reasoning traces they cannot inspect, risks losing nuance through internal translation, and discards knowledge that may be more naturally accessible in the target language. The paper asks how to optimize data composition and scheduling during SFT so that a model generalizes "in-language" (L2) reasoning — reasoning in the language of the prompt — to many languages at once, including ones never directly supervised.

Key Contributions

1. A data-mixing recipe for L2 reasoning. The authors systematically compare ways of combining English reasoning (ER), multilingual reasoning (MR), and multilingual non-reasoning (NR) data, and show that joint in-batch mixing beats sequential adaptation and weight merging for the accuracy/L2-rate trade-off.

2. Tiny Aya L2-Thinker, a 3.35B dual-mode model. Trained on top of the massively multilingual Tiny Aya base, it reasons in the prompt language on over 93% of traces across 60 benchmark languages and 6 tasks spanning math, commonsense, instruction following, open-ended generation, and cultural reasoning.

3. A demonstration that reasoning is a language-agnostic, transferable behavior. Broad language coverage plus a modest seed set (<5K translated traces per language) and readily available multilingual instruction data is enough to generalize L2 reasoning to held-out languages — no reasoning supervision needed in every target language.

4. Open releases. Model weights (Tiny Aya L2-Thinker, Tiny Aya En-Thinker, Tiny Aya Base-32K) and the multilingual reasoning dataset are released publicly to support further research on accessible in-language reasoning.

Main Findings

• Broad language coverage in reasoning supervision improves transfer rather than causing interference. Adding more languages to the MR mix keeps accuracy and L2 rates steady on covered languages while raising L2 rates on languages never supervised. The classical "curse of multilinguality" does not appear to apply to reasoning behavior, and a single jointly trained model outperforms per-region specialists.

• Multilingual non-reasoning instruction data is a cheap lever. Including ~4.9M multilingual NR samples (with empty thinking blocks) improves both accuracy and L2 reasoning rate on unseen languages and prevents forgetting of Tiny Aya's 67 languages, at far lower cost than translating reasoning traces.

• English reasoning data remains the necessary backbone. It supplies the core task-solving capability, especially for difficult math; removing it degrades performance sharply even when multilingual supervision is present.

• Mixing beats sequential adaptation and weight merging. How supervision is combined matters as much as what it contains; joint training gives a better accuracy/L2-rate trade-off than staged fine-tuning or merging separately trained models.

• In-language reasoning costs little accuracy. Switching from English-only to L2 reasoning drops other-language accuracy by at most 2–3 points on five of six benchmarks; the one real loss is on competition-level math (PolyMath: 11.1% vs. 18.6% for the English reasoner), which the authors attribute to the absence of an RL stage.

• Inference-time language forcing is fragile. Qwen3.5-4B with "thinking prefix" language forcing reaches high L2 rates but drops about 12 points in other-language accuracy on PolyMath and MIST-OEG, and its traces are far more repetitive. Tiny Aya L2-Thinker exceeds it on L2 rate for every benchmark and is more stable across languages.

• Trained L2 reasoning outperforms larger specialist models. At 3.35B, Tiny Aya L2-Thinker beats M-Thinker-7B (87.7–96.5% L2 rate) on five of six benchmarks and outperforms Magistral-Small-24B on all tasks by a wide margin, with far lower cross-language variance.

• Long traces correlate with doomlooping, not deeper reasoning. The 4-gram repetition score rises with the number of thinking tokens; Qwen3.5-4B sits at the high-token, high-repetition extreme, while Tiny Aya L2-Thinker is concise and low-repetition except on genuinely hard PolyMath.

• Low-resource languages are where the gap matters most. Across four resourcedness tiers, Tiny Aya L2-Thinker holds L2 rates of 94–99% while Magistral-Small-24B collapses from 72% to 5% and M-Thinker-7B drops from 99.6% to 71.6% with doubling trace length.

Methodology in Plain English

The authors start from Tiny Aya, a 3.35B multilingual base model, and extend its context window from 8K to 32K tokens using interleaved long- and short-context training. They then assemble three data sources:

• English reasoning (ER): ~1.7M samples from AM-Thinking and gpt-oss-120b, later augmented with longer-trace datasets for the final model. This teaches the model how to reason.
• Multilingual reasoning (MR): The same English thinking traces machine-translated into 44 diverse languages (capped at ~5K samples per language, chosen to cover different language families, scripts, and resource levels). This teaches the model which language to reason in.
• Multilingual non-reasoning (NR): ~4.9M ordinary instruction-following examples across 67 languages, included with empty thinking blocks so the model learns to answer directly when no reasoning is needed and retains general multilingual ability.

Rather than training separate models per language or adapting an English reasoner post-hoc, the authors mix all three sources within each training batch. This "dual-mode" setup lets the model learn to reason in English, reason in other languages, and follow plain instructions simultaneously. They evaluate on six benchmarks — MGSM and PolyMath (math), GlobalPIQA and Macaron-MCQ (commonsense and cultural reasoning), Marco-Bench-MIF (localized instruction following), and MIST-OEG (open-ended generation) — tracking two metrics per language: task accuracy and L2 reasoning rate (measured by a FastText language identifier, with GlotLID as a fallback), plus repetition rate and thinking-token length as efficiency proxies.

Why This Matters

Impact on research. The paper reframes multilingual reasoning from "a capability to be added per language" to "a behavior that transfers through data mixing," and provides controlled evidence that the curse of multilinguality does not apply to reasoning. It also challenges the assumption that L2 reasoning must come at a "multilinguality tax," offering a reproducible data recipe and open artifacts for others to build on.

Real-world applications:

• Healthcare and medical reasoning, where clinicians need to inspect and trust a model's intermediate reasoning in the local language rather than in English.
• Cultural and region-specific question answering, where locally embedded knowledge, terminology, and notions of harm are more reliably expressed in the target language.
• Low-resource deployment, where a 3.35B model that reasons natively in dozens of languages is far more practical than a 24B+ alternative.
• Localized instruction following and creative writing, where naturalness and coherence in the user's language matter as much as factual correctness.

Industry relevance. The results offer a concrete playbook for labs building multilingual products: invest in broad language coverage at small per-language volumes, keep English reasoning as the backbone, and mix in cheap multilingual instruction data instead of relying on inference-time prompting tricks. The open weights and dataset lower the barrier for startups and researchers without the budget for per-language reasoning data or large-scale RL.

Future Directions

• Does reinforcement learning close the math gap? The authors attribute their PolyMath deficit to the absence of an RL stage; testing whether RL refinement preserves L2 behavior while recovering competition-level math accuracy is a clear next step.

• Does the recipe hold at larger scales and in bigger MoE models? The 3.35B conclusions may not transfer directly to frontier-scale models with different data mixtures and tokenizers.

• How far can translation-quality filtering be pushed? The paper explicitly treats translated reasoning as a scarce, carefully filtered resource; whether scaling up translation further, or using stronger translator models, changes the generalization curve remains open.

• Are there limits to language-agnostic transfer? The paper shows L2 reasoning transfers to unseen languages, but the boundary of that transfer — for languages with no NR data at all, or for reasoning-intensive domains like code — is not yet mapped.

Target Audience

This paper is most useful to researchers and engineers working on multilingual LLMs, reasoning model training, and data-centric fine-tuning; to teams shipping multilingual products who need in-language reasoning without a large compute budget; and to anyone studying cross-lingual transfer, the "curse of multilinguality," or the accessibility of AI systems for non-English speakers. Readers with a basic understanding of supervised fine-tuning and chain-of-thought reasoning will get the most out of it.

Authors’ abstract

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.