Research map / Large language models / Agents, code and math
BOOSTAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models
Yuanhao Li, Hongbo Wang, Xiaotang Shang, Xunzhu Tang, Yiming Cao, Xuhong Chen
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 PPO with sequence-level reward model Rseq after SFT (ablation) | BoostAPR (Qwen2.5-Coder-32B) · SWE-bench Verified | pass@1 (%): 23.4 -> 38.3 | Improved |
| add execution-verified SFT with reasoning traces (ablation) | Qwen2.5-Coder-32B base · SWE-bench Verified | pass@1 (%): 17.8 -> 23.4 | Improved |
| add line-level credit allocator Rline to PPO with Rseq (ablation) | BoostAPR · SWE-bench Verified | pass@1 (%): 38.3 -> 40.7 | Improved |
| add line-level credit allocator Rline to PPO with Rseq (ablation) | BoostAPR · Defects4J v2.0 (Java transfer) | resolve rate (%): 19.2 -> 24.8 | Improved |
| add line-level credit allocator Rline to PPO with Rseq (ablation) | BoostAPR · QuixBugs | fixed bugs (%): 95.0 -> 95.0 | No clear effect |
| use Rseq only for reranking instead of PPO training (ablation) | SFT policy · SWE-bench Verified | pass@1 (%): 26.1 vs 38.3 with PPO | Hurt |
| use edit-line spans instead of token-uniform credit (ablation) | BoostAPR · SWE-bench Verified | resolve rate (%): 37.6 -> 40.7 | Improved |
| use edit-line spans instead of hunk-level credit (ablation) | BoostAPR · Defects4J v2.0 (Java transfer) | resolve rate (%): 21.3 -> 24.8 | Improved |
| use edit-line spans instead of statement-level credit (ablation) | BoostAPR · Defects4J v2.0 (Java transfer) | resolve rate (%): 22.1 -> 24.8 | Improved |
| score patches only (no bug context) in Rseq (ablation) | Rseq reward model · SWE-bench Verified | pass@1 (%): 38.9 -> 40.7 | Improved |
| train Rseq with hybrid regression plus preference objective (ablation) | Rseq reward model · SWE-bench Verified | pass@1 (%): 39.2 -> 40.7 | Improved |
| BoostAPR vs GRPO under same backbone and data | Qwen2.5-Coder-32B · SWE-bench Verified | resolve rate (%): 36.1 -> 40.7 | Improved |
Closest papers by meaning
- Less Is More: Adaptive Program Repair with Bug Localization and Preference Learning (AAAI 2025 · similarity 0.87)
- StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning (ICML 2026 · similarity 0.87)
- Code Repair with LLMs gives an Exploration-Exploitation Tradeoff (NeurIPS 2024 · similarity 0.86)
- Counterexample Guided Program Repair Using Zero-Shot Learning and MaxSAT-based Fault Localization (AAAI 2025 · similarity 0.86)
- AuPair: Golden Example Pairs for Code Repair (ICML 2025 · similarity 0.86)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.85)
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.
