Reproduce “Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaofeng Wu, Shiqi Wang, Boya Peng, Anuj Goyal, Melanie Kambadur, Sebastian Ruder, Yoon Kim, Chloe Bi
Researchers at Meta Superintelligence Labs and MIT introduced Parallel-SFT, a supervised fine-tuning strategy that uses synthetically generated parallel programs to enhance zero-shot cross-programming-language transfer for code Reinforcement Learning (RL). This method improves model generalization and aligns internal representations across diverse programming languages, leading to performance gains in unseen target languages.
Key insights
- Effective zero-shot cross-PL transfer in code RL is significantly reliant on a generalizable SFT initialization that can align representations across different programming languages.
- Explicitly incorporating parallel program signals during SFT fosters a functionality-centric understanding of code, enabling models to generalize based on execution equivalence rather than surface syntax.
- Analysis of internal model representations supports the "semantic hub hypothesis" for code, showing that Parallel-SFT induces stronger cross-PL representation alignment, particularly in intermediate model layers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.