Research map / Reinforcement learning / RL from feedback
GXPO: Group Cross-Lingual Relative Policy Optimization for Code Generation
Linzheng Chai, Jian Yang, Jiajun Wu, Ensheng Shi, Xianglong Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add monolingual advantage (ablation) | GXPO · ML-LCB v6 Avg (Qwen2.5-Coder-7B) | Pass@1: 20.9 -> 22.6 | Improved |
| add cross-lingual advantage (ablation) | GXPO · ML-LCB v6 Avg (Qwen2.5-Coder-7B) | Pass@1: 21.6 -> 22.6 | Improved |
| add multilingual group advantage (ablation) | GXPO · ML-LCB v6 Avg (Qwen2.5-Coder-7B) | Pass@1: 21.1 -> 22.6 | Improved |
| add monolingual advantage (ablation) | GXPO · ML-LCB v6 Low-resource PLs | Pass@1: 20.4 -> 22.9 | Improved |
| add cross-lingual advantage (ablation) | GXPO · ML-LCB v6 High-resource PLs | Pass@1: 21.3 -> 22.3 | Improved |
| add multilingual group advantage (ablation) | GXPO · ML-LCB v6 Low-resource PLs | Pass@1: 20.4 -> 22.9 | Improved |
| replace GRPO with GXPO | Qwen3-4B-Base · ML-LCB v6 Avg | Pass@1: 15.0 -> 19.1 | Improved |
| replace GRPO with GXPO | LLaMA-3.1-8B · ML-LCB v6 Avg | Pass@1: 11.6 -> 14.8 | Improved |
| replace GRPO with GXPO | Qwen2.5-Coder-7B-Instruct · ML-LCB v6 Avg | Pass@1: 21.8 -> 22.6 | Improved |
| replace GRPO with GXPO | Qwen2.5-Coder-7B-Instruct · ML-LCB v5 Avg | Pass@1: 16.6 -> 18.9 | Improved |
| replace GRPO with GXPO | LLaMA-3.1-8B · ML-LCB v6 Perl | Pass@1: 0.0 -> 14.4 | Improved |
| replace GRPO with GXPO | LLaMA-3.1-8B · ML-LCB v6 Low-resource vs High-resource gap | Pass@1: 10.2 -> 14.8 | Improved |
Closest papers by meaning
- On-Policy Optimization with Group Equivalent Preference for Multi-Programming Language Understanding (NeurIPS 2025 · similarity 0.91)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.89)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.88)
- Stronger-MAS: Multi-Agent Reinforcement Learning for Collaborative LLMs (ICLR 2026 · similarity 0.88)
- Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search (NeurIPS 2024 · similarity 0.87)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
