---
title: "Summary: 2026-04-22_17-58-36Z_Parallel_SFT_ImprovingZero_ShotCross_Programming_L"
date: 2026-04-22
tags: ['paper', 'research', 'ai']
---
# Summary: 2026-04-22_17-58-36Z_Parallel_SFT_ImprovingZero_ShotCross_Programming_L


**Source**: [Original Paper](http://arxiv.org/abs/2604.20835v1)
Saved: 2026-05-08 03:29
Source: 2026-04-22_17-58-36Z_Parallel_SFT_ImprovingZero_ShotCross_Programming_L.md
Model: None

---

## Summary
Parallel-SFT targets zero-shot transfer of code RL across programming languages. It uses parallel programs in supervised fine-tuning to make representations more functionality-centric, which improves transfer to unseen languages after RL.

## Semantic links
- [[concepts/papers/2026-06-11_15-09-32Z_TowardInstructions_as_Code_Understandingthe_summary.md|Summary: 2026-06-11_15-09-32Z_TowardInstructions_as_Code_UnderstandingtheImpacto.md]] — 3 title terms overlap; shared tags: ai, paper, research; 5 summary/topic terms overlap
- [[concepts/papers/2026-06-11_17-59-36Z_SpatialClaw_RethinkingActionInterfaceforAge_summary.md|Summary: 2026-06-11_17-59-36Z_SpatialClaw_RethinkingActionInterfaceforAgenticSpa.md]] — 3 title terms overlap; shared tags: ai, paper, research; 4 summary/topic terms overlap

## Key Takeaways
- RL trained on one source language can fail to transfer to other programming languages.
- Parallel-SFT injects functionally equivalent code in multiple languages during SFT.
- The resulting model generalizes better to unseen languages after RL.


## Context
The paper studies how programming skills learned in one language can be transferred to lower-resource languages.

## Implications
A better SFT initialization can make later RL more transferable across programming languages.

## Original Reference
- Title: Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL
- Authors: Zhaofeng Wu, Shiqi Wang, Boya Peng, Anuj Goyal, Melanie Kambadur, Sebastian Ruder, Yoon Kim, Chloe Bi
- URL: http://arxiv.org/abs/2604.20835v1
- Published: 2026-04-22T17:58:36Z
- Source file: /home/rich/wiki/ai-research/raw/papers/2026-04-22_17-58-36Z_Parallel_SFT_ImprovingZero_ShotCross_Programming_L.md

[[raw/papers/2026-04-22_17-58-36Z_Parallel_SFT_ImprovingZero_ShotCross_Programming_L.md]]

## Related Concepts

- [[concepts/software-development/software-development-hub.md|Software Development Hub]]
- [[concepts/search-retrieval/search-retrieval-hub.md|Search Retrieval Hub]]
- [[concepts/training-optimization/training-optimization-hub.md|Training Optimization Hub]]
- [[concepts/audio-speech/audio-speech-hub.md|Audio Speech Hub]]