Research map / Label-efficient and robust learning / Self-supervised and contrastive
Graph-based, Self-Supervised Program Repair from Diagnostic Feedback
Michihiro Yasunaga, Percy Liang
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 the program-feedback graph to the sequence model (ablation) | DrRepair · DeepFix raw test | full repair rate: 62.5 -> 66.4 | Improved |
| Add self-supervised pre-training (ablation) | Base + graph · DeepFix raw test | full repair rate: 66.4 -> 68.2 | Improved |
| Remove compiler messages (ablation) | Base repair model · DeepFix raw test | full repair rate: 62.5 -> 34.0 | Hurt |
| Add the program-feedback graph (ablation) | SPoC repair model · SPoC dev | single repair accuracy: 48.6 -> 53.0 | Improved |
| Add self-supervised pre-training (ablation) | SPoC Base + graph · SPoC dev | single repair accuracy: 53.0 -> 56.2 | Improved |
| Provide pseudocode as an additional input (ablation) | DrRepair · SPoC dev | single repair accuracy: 56.2 -> 68.0 | Improved |
| Use DrPerturb instead of Gupta+17 corruption (ablation) | Base repair model · DeepFix raw test | full repair rate: 50.5 -> 62.5 | Improved |
| Use DrPerturb instead of random token-drop corruption (ablation) | Base repair model · DeepFix raw test | full repair rate: 49.4 -> 62.5 | Improved |
| Replace the program-feedback graph with code-only edges (ablation) | Base + graph · DeepFix raw test | full repair rate: 66.4 -> 64.8 | Hurt |
| Replace the program-feedback graph with code-feedback edges only (ablation) | Base + graph · DeepFix raw test | full repair rate: 66.4 -> 64.9 | Hurt |
| Replace the program-feedback graph with self-attention (ablation) | Base repair model · DeepFix raw test | full repair rate: 66.4 -> 66.0 | Hurt |
| Add DrRepair to best-first program synthesis search | SPoC synthesis · SPoC TestP, B=100 | synthesis success rate: 34.2 -> 37.5 | Improved |
Closest papers by meaning
- Counterexample Guided Program Repair Using Zero-Shot Learning and MaxSAT-based Fault Localization (AAAI 2025 · similarity 0.90)
- Break-It-Fix-It: Unsupervised Learning for Program Repair (ICML 2021 · similarity 0.88)
- Is Self-Repair a Silver Bullet for Code Generation? (ICLR 2024 · similarity 0.87)
- AuPair: Golden Example Pairs for Code Repair (ICML 2025 · similarity 0.86)
- Less Is More: Adaptive Program Repair with Bug Localization and Preference Learning (AAAI 2025 · similarity 0.86)
- Code Repair with LLMs gives an Exploration-Exploitation Tradeoff (NeurIPS 2024 · 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.
