Crosslingual On-Policy Self-Distillation for Multilingual Reasoning
Summary
The paper proposes Crosslingual On-Policy Self-Distillation (COPSD), a method to transfer high-resource language reasoning capabilities to low-resource languages using a shared student-teacher architecture. Experiments across 17 African languages show significant improvements in mathematical reasoning and answer-format adherence, outperforming Group Relative Policy Optimization (GRPO).
View Cached Full Text
Cached at: 05/12/26, 10:51 AM
Paper page - Crosslingual On-Policy Self-Distillation for Multilingual Reasoning
Source: https://huggingface.co/papers/2605.09548
Abstract
COPSD transfers high-resource language model reasoning behavior to low-resource languages using self-distillation with crosslingual context, improving mathematical reasoning performance.
Large language models(LLMs) have achieved remarkable progress inmathematical reasoning, but this ability is not equally accessible across languages. Especiallylow-resource languagesexhibit much lower reasoning performance. To address this, we proposeCrosslingual On-Policy Self-Distillation(COPSD), which transfers a model’s own high-resource reasoning behavior tolow-resource languages. COPSD uses the same model as student and teacher: the student sees only the low-resource problem, while the teacher receives privileged crosslingual context, including the problem translation and reference solution in English. Training minimizes full-distributiontoken-level divergenceon the student’s own rollouts, providing dense supervision while avoiding the sparsity and instability of outcome-onlyreinforcement learning(RL). Experiments on 17 low-resource African languages show that COPSD consistently improves low-resourcemathematical reasoningacross model sizes and substantially outperforms Group RelativePolicy Optimization(GRPO). Further analyses show that COPSD improves answer-format adherence, strengthenstest-time scaling, and generalizes to hardermultilingual reasoning benchmarks, with especially large gains for lower-resource languages. We make our code and data available at: https://github.com/cisnlp/COPSD.
View arXiv pageView PDFGitHubAdd to collection
Get this paper in your agent:
hf papers read 2605\.09548
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2605.09548 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2605.09548 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2605.09548 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Efficient Multilingual Reasoning Transfer via Progressive Code-Switching
This paper introduces Progressive Code-Switching (PCS), a reinforcement learning approach with curriculum learning that gradually increases code-switching in LLMs to efficiently transfer multilingual reasoning capabilities.
Reasoning Compression with Mixed-Policy Distillation
This paper proposes Mixed-Policy Distillation (MPD), a framework that transfers concise reasoning behaviors from large teacher models to smaller student models, reducing token usage by up to 27.1% while improving performance.
dOPSD: On-Policy Self-Distillation for Diffusion Language Models
This paper introduces dOPSD, an on-policy self-distillation method for diffusion language models that leverages internal denoising trajectories to improve mathematical reasoning and code generation.
Respecting Self-Uncertainty in On-Policy Self-Distillation for Efficient LLM Reasoning
The paper proposes EGRSD and CL-EGRSD, on-policy self-distillation methods that weight token-level supervision by teacher entropy to improve reasoning accuracy-length tradeoff in LLMs, evaluated on Qwen3-4B and Qwen3-8B.
OPRD: On-Policy Representation Distillation
OPRD proposes a new knowledge distillation method that aligns student and teacher hidden states across layers during on-policy rollouts, eliminating sampling variance from token-space KL estimation. Empirically, OPRD outperforms output-space baselines on math reasoning benchmarks (AIME 2024/2025, AIMO) while being 1.44x faster and using 54% less memory.