Multilingual Reasoning Gym: Multilingual Scaling of Procedural Reasoning Environments

This paper introduces the Multilingual Reasoning Gym, a procedurally generated framework that extends the original Reasoning Gym to 14 languages with native-speaker validation, enabling the scalable creation of parallel, verifiable reasoning problems for training and evaluating multilingual models.

The Gist

Imagine you are trying to teach a robot how to solve puzzles. For a long time, you've only been able to teach it using English puzzles. The robot gets really good at English math and logic, but if you switch the puzzle to Japanese, German, or Swahili, the robot gets confused because it hasn't practiced those languages.

This paper introduces a new tool called the Multilingual Reasoning Gym. Think of it as a giant, magical puzzle factory that can instantly create millions of unique brain-teasers in 14 different languages.

Here is how it works, broken down with some everyday analogies:

1. The Problem: The "Static Library" vs. The "Infinite Factory"

Before this, researchers had to use "Static Libraries" (like a bookshelf). They would take a book of math problems, translate the whole book into French, then translate the whole book into Spanish.

• The Catch: There are only so many pages in a book. Once the robot has read all the translated books, it memorizes the answers instead of learning how to think. Also, some languages are "underserved," meaning there are very few translated books for them.

The Solution: The Multilingual Reasoning Gym is an Infinite Factory.
Instead of writing out every single puzzle, the researchers wrote a set of master templates (like a recipe).

• The Recipe: "Take two numbers, add them, and ask for the sum."
• The Magic: The factory can use this one recipe to bake 1,000,000 different cakes (puzzles) instantly. It can bake them in English, then immediately bake 1,000,000 different cakes in Japanese, using the exact same recipe logic.

2. The Challenge: It's Not Just "Google Translate"

You can't just run a machine translator on these recipes. If you translate a math problem literally, it often sounds weird or breaks the rules of the language.

• The Analogy: Imagine a recipe that says, "Add three s to the end of the word." In English, that works for "cat" $\rightarrow$ "cats." But in German or Japanese, you can't just add an "s." The word changes completely.
• The Fix: The team didn't just translate the words; they re-wrote the instructions for each language.
  • They made sure Japanese used the correct punctuation (like full-width commas).
  • They swapped English math terms for the specific terms used in German schools.
  • They even changed how they asked for answers so the robot wouldn't get confused by grammar rules that don't exist in English.

They used a team of human experts (native speakers) to taste-test the recipes, ensuring the puzzles sounded natural and fair, not like a robot wrote them.

3. The Result: A Global Training Ground

Now, researchers can train their AI models using this gym.

• Fair Play: Because the factory uses the same "seed" (the same starting point) for all languages, they can generate a puzzle in English and the exact same puzzle in Korean at the same time. This lets them test if the AI is actually smart, or if it's just good at English.
• Adjustable Difficulty: Just like a video game, you can turn the dial from "Easy" (for a beginner robot) to "Hard" (for a pro robot) instantly, in any language.

Why Does This Matter?

Think of AI models as students.

• Before: The student only studied in an English classroom. They were great at English math but failed when the teacher switched to Spanish.
• Now: The Multilingual Reasoning Gym puts the student in a classroom where they can practice math, logic, and coding in 14 different languages simultaneously. It ensures that when the AI talks to a user in Swahili or Thai, it's just as smart as when it talks to a user in English.

In short: This paper gives us a machine that can generate infinite, high-quality logic puzzles in many languages, helping us build AI that is truly smart and fair for everyone, not just English speakers.

Technical Summary

Here is a detailed technical summary of the paper "Multilingual Reasoning Gym: Multilingual Scaling of Procedural Reasoning Environments."

1. Problem Statement

While Reinforcement Learning with Verifiable Rewards (RLVR) has proven highly effective for training Large Language Models (LLMs) on reasoning tasks (math, logic, algorithms), existing frameworks like the original Reasoning Gym are limited to English. This creates a significant bottleneck for multilingual research and deployment because:

• Static Datasets are Limited: Existing multilingual benchmarks (e.g., MGSM, MMATH) are static, finite, and prone to data contamination or overfitting.
• Lack of Parallelism: Static datasets rarely offer perfectly parallel instances across languages, hindering cross-lingual analysis.
• Translation Challenges: Simply translating static datasets does not address the need for procedural generation, which allows for infinite data generation and adjustable difficulty curves essential for curriculum learning.
• Linguistic Nuance: Direct translation of templates often fails to account for morphological differences, notation conventions, and cultural context, leading to unnatural or logically flawed problem instances in non-English languages.

2. Methodology

The authors present the Multilingual Reasoning Gym, an extension of the original Reasoning Gym that procedurally generates verifiable reasoning problems across 14 languages. The methodology involves a hybrid human-LLM workflow and significant architectural adaptations.

A. Scope and Languages

The framework supports 94 tasks across 14 languages:

• High-resource: English, Chinese, German, Spanish, French, Japanese, Brazilian Portuguese, Russian, Korean, Italian.
• Low-resource/Underserved: Thai, Bengali, Telugu, Swahili.

B. Technical Architecture & Adaptations

To move from English-only to multilingual procedural generation, the authors implemented several key technical changes:

1. Template Extraction: All hard-coded English strings were extracted from the task logic into JSON template files. This decouples the generation logic from the language, allowing new languages to be added simply by providing a new JSON file.
2. Task Filtering & Adaptation:
   • Skipped Tasks: 10 tasks were excluded entirely because they rely on English-specific properties (e.g., word_ladder requires English vocabulary; figlet_fonts relies on ASCII art incompatible with non-Latin scripts).
   • Adapted Tasks: 14 tasks were modified to handle linguistic differences. For example, pluralization logic (common in English) was restructured to avoid morphological errors in languages like German or Russian (e.g., changing "tiger{suffix}" to "tiger: {number}").
   • Hybrid Content: 6 tasks retain English source material (e.g., reversing English sentences) but are translated with clear instructions that the input data is English.
3. Translation Workflow:
   • LLM Refinement: Initial translations were generated using Claude Sonnet 4, followed by an iterative refinement loop where translations were graded against rubrics and polished.
   • Native Speaker Validation: Two human native speakers validated samples for 10 major languages (Chinese, German, French, Spanish, Italian, Russian, Thai, Japanese, Korean, Portuguese).
   • Underserved Languages: For Bengali, Telugu, and Swahili, translations were provided without native speaker validation to increase coverage, though the authors note this may impact fluency consistency.
   • Verification: The verification logic (code) remains identical across languages; only the surface form (prompts, answers, punctuation) changes.

3. Key Contributions

• First Large-Scale Multilingual RLVR Environment: The paper introduces the first framework capable of generating cross-lingually parallel reasoning problems at an unlimited scale with verifiable rewards.
• Procedural Generation at Scale: Unlike static datasets, this system allows for the generation of infinite unique problem instances with adjustable difficulty, enabling curriculum learning tailored to specific model capabilities in different languages.
• Hybrid Translation Pipeline: A robust pipeline combining LLMs, iterative refinement, and native speaker validation to ensure mathematical and logical correctness alongside linguistic naturalness.
• Open Release: The authors released the code and data (94 tasks, 14 languages) to facilitate research into multilingual reasoning models.

4. Experimental Results

The authors evaluated various open-weight models (including Gemma-3, Qwen3, SmolLM3, and gpt-oss) on the dataset across three difficulty levels (25th, 75th percentiles, and default).

• Performance Trends:
  • Capacity Scaling: Performance consistently increased with model size across all languages.
  • Language Disparity: English (en) consistently achieved the highest scores. Non-English languages showed varying performance gaps, with some (like Japanese and Korean) performing relatively well, while others (like Telugu and Swahili) showed lower scores, partly attributed to the lack of native speaker validation for these specific languages.
  • Difficulty Sensitivity: The system successfully demonstrated a wide range of difficulty. Performance dropped significantly (up to 31 percentage points) when moving from the easiest configuration to the hardest (75th percentile), proving the curriculum's effectiveness.
• Language Consistency:
  • A critical finding was "multilingual confusion." Models often reasoned in English even when prompted in other languages (e.g., Qwen3 reasoning in English for German prompts).
  • Furthermore, models sometimes solved the problem correctly but failed to translate the final answer back to the target language (e.g., answering "green" instead of "grün"), leading to verification failures.

5. Significance and Impact

• Enabling Multilingual RLVR: The paper provides the necessary infrastructure to train and evaluate reasoning models in languages other than English using the powerful RLVR paradigm, which was previously restricted to English.
• Mitigating Data Scarcity: By using procedural generation, the work solves the data scarcity problem for low-resource languages, offering a theoretically infinite supply of training data.
• Benchmarking Tool: It serves as a rigorous benchmark for testing the cross-lingual generalization of LLMs, specifically highlighting gaps in reasoning consistency and answer formatting across languages.
• Future Research Direction: The identified limitations (e.g., models reasoning in English despite non-English prompts) highlight a new frontier for research: training models to maintain language-consistent reasoning chains and output formatting.

In conclusion, the Multilingual Reasoning Gym bridges the gap between the scalability of procedural generation and the necessity of multilingual reasoning, providing a foundational tool for the next generation of globally capable AI models.