Research map / Large language models / Instruction and fine-tuning
ReCode: Updating Code API Knowledge with Reinforcement Learning
Haoze Wu, Yunzhi Yao, Wenhao Yu, Ningyu Zhang
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 ReCode (DAPO) RL training | Qwen2.5-Coder-7B-Instruct · CodeUpdateArena | Pass@1: 67.3 -> 78.7 | Improved |
| add ReCode (GRPO) RL training | Qwen2.5-Coder-7B-Instruct · CodeUpdateArena | Pass@1: 67.3 -> 74.6 | Improved |
| add ReCode (DAPO) RL training | DS-v1.5-Coder-7B-Instruct · CodeUpdateArena | Pass@5: 72.5 -> 78.2 | Improved |
| replace SFT with ReCode (DAPO) | DS-v1.5-Coder-7B-Instruct · CodeUpdateArena | Pass@1: 53.4 -> 63.6 | Improved |
| replace SFT with ReCode (DAPO) to preserve general ability | Qwen2.5-Coder-7B-Instruct · HumanEval+ | Pass@1: 70.2 -> 81.7 | Improved |
| add syntax checking (AST) to ES reward (ablation) | ReCode GRPO · CodeUpdateArena reward ablation | Pass@1: +5.4 -> +7.4 | Improved |
| add syntax checking (AST) to EM reward (ablation) | ReCode GRPO · CodeUpdateArena reward ablation | Pass@1: -1.2 -> +1.1 | Improved |
| use edit-similarity (ES*) instead of exact match (EM*) reward (ablation) | ReCode GRPO · CodeUpdateArena reward ablation | Pass@5: +2.0 -> +8.0 | Improved |
| add correctness reward to format-only reward (ablation) | ReCode GRPO · CodeUpdateArena reward ablation | Pass@1: -2.3 -> +7.4 | Improved |
| add ReCode to model on multiple API updates | Qwen2.5-Coder-7B-Instruct · 20 multi-API-update questions | Pass@1: 35 -> 60 | Improved |
| fix CodeUpdateArena test-case errors | DS-Coder-1.5 · CodeUpdateArena | Pass@1: 46.8 -> 59.1 | Improved |
| add ReCode (DAPO) on DS-Coder for general ability | DS-v1.5-Coder-7B-Instruct · HumanEval+ | Pass@1: 71.3 -> 68.9 | Hurt |
Closest papers by meaning
- CodeSync: Synchronizing Large Language Models with Dynamic Code Evolution at Scale (ICML 2025 · similarity 0.89)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.88)
- LibEvoBench: Probing Temporal Knowledge Stratification in Code Generation Models (NeurIPS 2026 · similarity 0.88)
- SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution (NeurIPS 2025 · similarity 0.87)
- R1-Code-Interpreter: LLMs Reason with Code via Supervised and Multi-stage Reinforcement Learning (ICLR 2026 · similarity 0.87)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · 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.
