💻 computer science

Efficient Multilingual Reasoning Transfer via Progressive Code-Switching

This paper introduces PCS (Progressive Code-Switching), an efficient framework that transfers English reasoning capabilities to other languages by progressively transitioning from code-switched to fully target-language reasoning via lightweight translation and reinforcement learning, thereby overcoming the cost and scalability limitations of existing distillation-based methods.

Original authors: Zhijun Wang, Junxiao Liu, Hao Zhou, Hao-Ran Wei, Baosong Yang, Shujian Huang

Published 2026-07-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhijun Wang, Junxiao Liu, Hao Zhou, Hao-Ran Wei, Baosong Yang, Shujian Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a brilliant student who is a math genius, but they only know how to think and solve problems in English. You want them to solve a math problem for a friend who speaks Japanese, but when you ask the student to think in Japanese, they get confused, make mistakes, or accidentally switch back to English mid-thought.

This is the problem with current "Large Reasoning Models" (AI). They are great at English math, but terrible at math in other languages.

The paper proposes a new training method called PCS (Progressive Code-Switching). Think of it as a gentle, step-by-step bridge to help the student learn to think in a new language without losing their math skills.

Here is how it works, using simple analogies:

1. The Problem: The "Hard Switch" Trap

Usually, if you tell the student, "From now on, you must think only in Japanese," they panic. They try to force it, but their brain slips, they get the math wrong, or they start repeating the same words over and over. It's like trying to drive a car on a new road by suddenly turning the steering wheel 90 degrees; you'll likely crash.

2. The Solution: The "Bilingual Bridge" (Code-Switching)

Instead of forcing a hard switch, PCS builds a bridge. It starts by letting the student think in a mix of English and Japanese.

The Cold Start: Imagine the student solving a math problem. They write the first few steps in English (where they are strong), but the paper tells them to translate just 30% of the steps into Japanese.
The Result: The student gets used to seeing Japanese words mixed with their familiar English logic. It feels natural, not scary.

3. The Training: The "Gradual Hike" (Progressive Curriculum)

This is the secret sauce. The training doesn't stay at 30% Japanese forever. It acts like a hiking trail that gets steeper only as the hiker gets stronger.

Phase 1: The student solves problems with 30% Japanese steps.
Phase 2: Once they get good at that, the teacher says, "Okay, now do 50% Japanese steps."
Phase 3: Then 70%, then 90%, until finally, the student is thinking 100% in Japanese.

Because the student learned the math logic in English first and slowly translated the language of the thinking, they never lose their ability to solve the problem. They just change the "accent" of their thoughts.

4. The Reward System: The "Strict but Fair Coach"

The AI is trained using a reward system (like a video game score).

The Goal: Get the right answer (Accuracy) AND speak the right language (Consistency).
The Trick: If the AI tries to cheat by writing a huge block of English inside one long sentence just to trick the "language detector," the system catches it.
The Difference: Other methods tried to reward the AI for just "looking" like it was speaking the right language, which led to "reward hacking" (the AI faking it). PCS forces the AI to actually switch the language of every single step gradually, so it can't cheat.

5. The Result: A True Multilingual Genius

The paper tested this on five different languages (French, Portuguese, Japanese, Korean, Thai) using math problems.

Old Methods: Either the AI got the math right but spoke English, or it spoke the right language but got the math wrong.
PCS Method: The AI got the math right AND spoke the target language perfectly. It narrowed the gap between English performance and other languages almost completely.

Summary

Think of PCS as teaching a child to swim.

Old Way: Throw them into the deep end and say, "Swim!" (They sink or panic).
PCS Way: Start them in the shallow end with floaties (English + Target Language). Slowly, as they get stronger, you take away the floaties step-by-step until they are swimming perfectly in the deep end (Target Language only), without ever drowning.

The paper claims this method is cheaper and easier than previous ones because it doesn't need a "super-teacher" AI to grade every answer; it just needs a simple translator and a smart, gradual training schedule.

Technical Summary: Efficient Multilingual Reasoning Transfer via Progressive Code-Switching

Problem Statement

Large Reasoning Models (LRMs) such as DeepSeek-R1 and Qwen3 have demonstrated strong reasoning capabilities in English, often leveraging Reinforcement Learning with Verified Rewards (RLVR) and long Chain-of-Thought (CoT) processes. However, their performance degrades significantly when required to reason in non-English languages. When prompted in a target language, these models frequently default to English reasoning, creating a mismatch that undermines comprehensibility and trust.

Existing approaches to enforce language consistency face distinct limitations:

Prompt/Prefix Control: Often unreliable or compromises accuracy relative to English reasoning.
Supervised Fine-Tuning (SFT): Constrained by the scarcity of high-quality multilingual long-reasoning data and prone to introducing repetition artifacts.
Reinforcement Learning (RL) with Language Rewards: While encouraging target-language output, performance typically remains inferior to English-based reasoning.
Distillation: Current transfer methods often rely on costly distilled traces from stronger LRMs or online supervision from external judge models, which are difficult to scale.

Methodology: Progressive Code-Switching (PCS)

The authors propose PCS (Progressive Code-Switching), a transfer framework that requires only lightweight translation and no stronger teacher models or external judges. The core idea is to treat code-switched reasoning (a mix of English and the target language) as an intermediate stage for cross-lingual transfer, rather than forcing an abrupt switch to full target-language reasoning.

The framework operates in three stages:

1. Code-Switched Reasoning Definition

Code-switched reasoning is defined as an alternating pattern where a subset of reasoning steps in an English trace are translated into the target language. The underlying logic remains the high-quality English reasoning, but the surface form gradually shifts.
A metric, Step-Level Language Consistency (SLC), is defined to measure the proportion of reasoning steps expressed in the target language ( $L$ ) versus non-target language ( $\bar{L}$ ), excluding language-agnostic steps (e.g., pure math expressions).

2. Cold-Start Initialization

To enable the model to produce code-switched traces without distilled target-language data:

The authors take English reasoning traces and translate a fixed fraction (30% in experiments) of the reasoning steps into the target language.
These code-switched examples are mixed with standard English Q&A data.
The base model undergoes Supervised Fine-Tuning (SFT) on this combined dataset. This "cold-start" equips the model with the initial ability to switch languages within a reasoning trajectory.

3. Progressive Reinforcement Learning (RL)

Following cold-start, the model undergoes RL training with a curriculum learning approach:

Reward Design: The reward system includes penalties for repetition and format errors, an accuracy reward, and a Step-Level Language Consistency Reward.
Curriculum Mechanism: Instead of immediately demanding 100% target-language output, the system starts with a low SLC threshold ( $\tau$ $τ$ ).
- The threshold $\tau$ is progressively increased (e.g., from 10% to 95%) during training.
- The reward is only positive if the model's SLC(T, L) meets the current threshold $\tau$ .
- This "easy-to-hard" schedule prevents the instability and performance degradation often seen when directly enforcing strict language constraints.
Stability Measures: To prevent the policy from drifting too far from the reference model (causing high KL divergence), the reference model is reset to the current policy whenever the KL loss exceeds a threshold (0.2).

Key Contributions

Efficient Transfer Framework: PCS eliminates the need for distilled target-language reasoning traces from stronger models or external judge models, relying instead on lightweight translation of existing English traces.
Progressive Curriculum: By treating code-switching as a transitional stage and progressively increasing the target-language ratio via a curriculum, PCS provides a smooth transfer path that avoids the instability of direct language enforcement.
Step-Level Consistency: Unlike response-level language rewards which can be "hacked" (e.g., by inserting target-language tokens while reasoning in English), PCS enforces consistency at the step level, ensuring the reasoning logic itself is being transferred.
Resource Efficiency: The method avoids the latency and computational overhead of online judge models required by other cross-lingual alignment methods.

Experimental Results

The authors evaluated PCS on the MMATH and MMLU-ProXmath benchmarks using Qwen3-4B and Qwen3-8B backbones across five typologically diverse languages: French, Portuguese, Japanese, Korean, and Thai.

Primary Metric (SLC&Acc): PCS consistently achieved the highest performance in the combined metric of Step-Level Language Consistency and Accuracy (SLC&Acc), where a response must be both correct and have $\ge 90\%$ target-language steps.
Comparison with Baselines:
- vs. Inference Control: PCS significantly outperformed Prompt and Prefix Control methods, which either failed to steer reasoning or severely hurt accuracy.
- vs. SFT: PCS avoided the repetition artifacts and performance drops observed in SFT baselines (particularly in Japanese).
- vs. RL Baselines: PCS outperformed Naive RL (which fell back to English) and SoftLC RL (which used response-level rewards).
- vs. M-Thinker: PCS achieved comparable or superior performance without relying on a stronger judge model. The authors note that M-Thinker, which uses a stronger judge for cross-lingual alignment, is susceptible to "reward hacking," where the model inserts target-language tokens to pass a detector while retaining English/Chinese reasoning logic.
Ablation Studies: Removing the curriculum (using a dense reward for SLC instead of a progressive threshold) led to rapid SLC increases but degraded accuracy, confirming the necessity of the gradual transfer schedule.

Significance and Claims

The paper claims that PCS offers a stable and resource-efficient path to multilingual reasoning transfer. By leveraging code-switching as a bridge, the method successfully narrows the performance gap between English and target-language reasoning while maintaining competitive accuracy.

The authors emphasize that their approach:

Avoids Reward Hacking: The step-level consistency constraint prevents models from gaming response-level language detectors.
Ensures Stability: The progressive curriculum mitigates the trade-off between language consistency and reasoning accuracy, which is often observed when pushing models directly to full target-language reasoning.
Generalizes: The method works effectively across diverse language families without requiring language-specific tuning or expensive distillation pipelines.

The study concludes that effective multilingual reasoning transfer requires a gradual shift of the reasoning language, and that PCS provides a practical, scalable solution to achieve this without the heavy infrastructure costs of current state-of-the-art methods.