Research map / Large language models / Alignment and preferences
AP2O-Coder: Adaptively Progressive Preference Optimization for Reducing Compilation and Runtime Errors in LLM-Generated Code
Jianqing Zhang, Wei Xia, Hande Dong, Qiang Lin, Jian Cao
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 adaptive error replay (ablation) | AP2O-Coder (L2H), Qwen2.5-Coder-7B · EvalPlus MBPP | pass@1: 82.7 -> 84.9 | Improved |
| add adaptive error replay (ablation) | AP2O-Coder (H2L), Qwen2.5-Coder-7B · EvalPlus MBPP | pass@1: 82.1 -> 85.4 | Improved |
| add adaptive error replay (ablation) | AP2O-Coder (L2H), Qwen2.5-7B · EvalPlus MBPP | pass@1: 81.6 -> 82.9 | Improved |
| add adaptive error replay (ablation) | AP2O-Coder (H2L), Qwen2.5-7B · EvalPlus MBPP | pass@1: 81.0 -> 82.2 | Improved |
| add adaptive error replay (ablation) | AP2O-Coder (L2H and H2L), Qwen2.5-Coder-7B · LiveCodeBench | pass@1: 18.4 -> 18.8 (L2H), 18.5 -> 19.0 (H2L) | Improved |
| add adaptive error replay (ablation) | AP2O-Coder (L2H and H2L), Qwen2.5-7B · LiveCodeBench | pass@1: 14.2 -> 15.1 (L2H), 13.6 -> 14.4 (H2L) | Improved |
| progressive error-type ordering plus replay instead of dynamic sampling (Dyn-DPO) (ablation) | AP2O-Coder (H2L), Qwen2.5-Coder-7B · EvalPlus MBPP | pass@1: 83.7 -> 85.4 | Improved |
| use AP2O-Coder (H2L) instead of DPO-Coder | Qwen2.5-Coder-7B · EvalPlus MBPP | pass@1: 83.5 -> 85.4 | Improved |
| use L2H instead of H2L ordering (ablation) | AP2O-Coder, Qwen2.5-Coder-0.5B · EvalPlus HumanEval | pass@1: 56.5 -> 61.5 | Improved |
| use H2L instead of L2H ordering (ablation) | AP2O-Coder, CodeLlama-34B · EvalPlus HumanEval | pass@1: 47.9 -> 49.6 | Improved |
| use H2L instead of L2H ordering (ablation) | AP2O-Coder, Qwen2.5-Coder-0.5B · LiveCodeBench v6 | pass@1: 3.3 -> 3.2 | No clear effect |
| curriculum ordering of DPO data (Curri-DPO-Coder) instead of DPO (ablation) | Qwen2.5-Coder-7B · EvalPlus MBPP | pass@1: 83.5 -> 81.7 | Hurt |
Closest papers by meaning
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.88)
- EffiBench-X: A Multi-Language Benchmark for Measuring Efficiency of LLM-Generated Code (NeurIPS 2025 · similarity 0.87)
- Mercury: A Code Efficiency Benchmark for Code Large Language Models (NeurIPS 2024 · similarity 0.87)
- EffiLearner: Enhancing Efficiency of Generated Code via Self-Optimization (NeurIPS 2024 · similarity 0.86)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.86)
- NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits (ICML 2025 · similarity 0.86)
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.
