Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL

Zhaofeng Wu, Shiqi Wang, Boya Peng, Anuj Goyal, Melanie Kambadur et al.|April 22, 2026arXiv

Key Takeaway

Training code models on parallel implementations of the same program across multiple languages creates a more language-agnostic understanding of coding logic, enabling better zero-shot transfer to new programming languages.

Summary

This paper tackles a practical problem: AI models trained to write code in popular languages like Python often perform poorly in less common languages. The researchers propose Parallel-SFT, a training method that uses code examples showing the same program written in multiple languages side-by-side.

training

Key Terms

supervised-fine-tuning zero-shot-generalization reinforcement-learning latent-space