Research map / Large language models / Agents, code and math
From Historical Patches to Repair Plans: Outcome-Conditioned Reasoning for Repository-Level Program Repair
Chenglin Li, Yisen Xu, Zehao Wang, Shin Hwei Tan, Tse-Hsun 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 the Exemplar Guardian filter (ablation) | ConRAD · SWE-Bench Lite | Pass@1 (%): 33.7 -> 37.7 (GPT-4o); 37.7 -> 40.0 (GPT-5); 30.0 -> 35.6 (DeepSeek-V3) | Improved |
| use backward reasoning distillation instead of MCTS forward exploration (ablation) | ConRAD · SWE-Bench Lite (100 samples) | Pass@1 (%): 28.0 -> 41.0 | Improved |
| MCTS forward search uses more compute than BRD (ablation) | ConRAD · SWE-Bench Lite (100 samples) | LLM calls and tokens: 9.71x more LLM calls, 2.47x more tokens | Hurt |
| add distilled reasoning plan on top of exemplar and patch injection (ablation) | ConRAD · SWE-Bench Lite (100 samples) | Pass@1 (%): 32.0 -> 41.0 | Improved |
| inject raw exemplar issue and ground-truth patch without a plan (ablation) | Agentless · SWE-Bench Lite (100 samples) | Pass@1 (%): 34.0 -> 32.0 | Hurt |
| use top-1 exemplar rather than top-3 exemplar (ablation) | ConRAD · SWE-Bench Lite (100 samples) | Pass@1 (%): 39.0 -> 45.0 | Improved |
| add ConRAD to Agentless | Agentless · SWE-Bench Lite | Pass@1 (%): 27.3 -> 37.7 (GPT-4o) | Improved |
| add ConRAD to Agentless with GPT-5 | Agentless · SWE-Bench Lite | Pass@1 (%): 29.7 -> 40.0 | Improved |
| add ConRAD to Agentless with DeepSeek-V3 | Agentless · SWE-Bench Lite | Pass@1 (%): 27.0 -> 35.6 | Improved |
| add ConRAD to mini-SWE-agent | mini-SWE-agent · SWE-Bench Lite (100 samples) | Pass@1 (%): 48.0% -> 58.0% | Improved |
| add ConRAD to Agentless with GPT-5 | Agentless · SWE-Bench Lite | file localization accuracy (%): 85.0% -> 86.3% | No clear effect |
| add ConRAD to Agentless with GPT-4o | Agentless · SWE-Bench Lite | function localization accuracy (%): 51.0% -> 59.3% | Improved |
Closest papers by meaning
- RepoMirage: Probing Repository Context Reasoning in Code Agents with Perturbations (NeurIPS 2026 · similarity 0.87)
- Counterexample Guided Program Repair Using Zero-Shot Learning and MaxSAT-based Fault Localization (AAAI 2025 · similarity 0.87)
- 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 0.86)
- AuPair: Golden Example Pairs for Code Repair (ICML 2025 · similarity 0.85)
- Build-and-Find: An Effort-Aware Protocol for Evaluating Agent-Managed Codebases (NeurIPS 2026 · similarity 0.84)
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.
