Back to Feed
Reasoning / Training & Fine-Tuning

Improving Multilingual Reasoning in Language Models

Original: RP-OPSD: Reasoning-Pivot-Guided On-Policy Self-Distillation for Multilingual Reasoning Transfer

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • RP-OPSD improves reasoning performance in low-resource languages compared to the COPSD baseline.
  • The method uses a Reasoning-Pivot Transfer gate to decide when to apply privileged information versus keeping target-language structure.
  • On AfriMGSM, the Qwen3-4B model achieved a pass@12 score of 26.83.
  • The approach effectively balances English-based reasoning transfer with local language preservation.

Summary & Methodology Analysis

The RP-OPSD methodology addresses the gap in multilingual reasoning by using an on-policy self-distillation framework, which is a process where a model learns by distilling knowledge from its own previous generations. The system constructs rollouts in the target language and compares two teacher views: one conditioned on an English reference trace and an ablated view that lacks this trace. By calculating the KL divergence (a statistical measure of how one probability distribution differs from another) between these views, the system generates a Reasoning-Pivot Transfer gate. This gate acts as a dynamic router that determines whether the model should follow the privileged English-based reasoning path or maintain the target language's natural realization.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main problem this paper solves?

It addresses the difficulty of transferring strong reasoning capabilities from high-resource English models to low-resource linguistic contexts.

Q2. How does the proposed method improve performance?

It uses a gate mechanism to selectively apply privileged information from English traces only when necessary, preserving the target language's structure elsewhere.

Q3. Did this approach perform better than previous methods?

Yes, it outperformed the COPSD baseline on the AfriMGSM benchmark by 2.37 points for Qwen3-1.7B and 5.20 points for Qwen3-4B.

Q4. What is a Reasoning-Pivot Transfer gate?

It is a computed value based on the KL divergence between a teacher view with an English reference trace and an ablated teacher view, used to route distillation.

Q5. What are the specific performance results on the PolyMath benchmark?

RP-OPSD achieved average DW-ACC of 17.97 for Qwen3-1.7B and 31.87 for Qwen3-4B.

Q6. What are the limitations of the reported performance?

The gains on PolyMath were more modest than on AfriMGSM, likely due to the DW-ACC metric weighting difficult problems that are inherently harder to solve.

Q7. Which models were used for evaluation?

The paper evaluated Qwen3-1.7B and Qwen3-4B models.

Q8. What does the objective function consist of?

It minimizes a combined objective of a pivot-transfer KL term and a language-anchoring KL term.

Q9. Does the paper specify hardware requirements for this method?

The paper does not specify hardware requirements.