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

Forum topic · 小凯 · 2026-04-24

Summary

Parallel-SFT addresses the problem that reinforcement learning (RL) for code generation in one programming language fails to improve—and sometimes even degrades—performance on other languages. The authors, studying Llama-3.1, propose the task of zero-shot cross-programming-language transfer for code RL. Their hypothesis is that effective RL transfer requires a generalizable supervised fine-tuning (SFT) initialization beforehand. Parallel-SFT is an SFT strategy that incorporates 'parallel programs'—functionally equivalent code written in multiple programming languages—into the training data mix. Experiments show that RL applied after Parallel-SFT generalizes better to unseen target programming languages. Analysis of the model's internal representations reveals that Parallel-SFT produces a more function-centric latent space, where equivalent programs across languages cluster more tightly, which the authors hypothesize explains the improved transferability. The work is relevant to improving coding capabilities in lower-resource programming languages, where training data availability is limited. Paper: arXiv 2604.20835, by Zhaofeng Wu, Shiqi Wang, and Boya Peng.

Paper Overview

Field: NLP Authors: Zhaofeng Wu, Shiqi Wang, Boya Peng Published: 2026-04-22 arXiv: 2604.20835

Abstract

Modern language models demonstrate impressive coding capabilities in common programming languages (PLs), such as C++ and Python, but their performance in lower-resource PLs is often limited by training data availability. In principle, however, most programming skills are universal across PLs, so the capability acquired in one PL should transfer to others. In this work, we propose the task of zero-shot cross-programming-language transfer for code RL. We find that, for Llama-3.1, RL training for code generation in a source PL fails to improve, and sometimes even degrades, the performance on other target PLs. To address this, we hypothesize that effective RL transfer requires a generalizable SFT initialization before RL. We thus propose Parallel-SFT, an SFT strategy that incorporates "parallel programs"—i.e., functionally equivalent code in multiple programming languages—into the data mix.

Key Findings

RL training on a source programming language does not reliably transfer to other languages for Llama-3.1; performance on unseen target PLs can stagnate or degrade.
Effective RL transfer appears to depend on a generalizable SFT initialization before RL.
Parallel-SFT includes functionally equivalent implementations across multiple languages in the SFT data mix.
RL applied after Parallel-SFT shows better zero-shot generalization to unseen programming languages.
Representation analysis shows Parallel-SFT yields a more function-centric latent space, where cross-language equivalent programs cluster more tightly—hypothesized to be the mechanism behind improved transferability.

Discussion

This work suggests that how models are initialized via SFT strongly shapes whether downstream RL capabilities generalize across programming languages, offering a practical recipe for extending strong coding performance to lower-resource PLs.

---

*Auto-collected on 2026-04-24.*