BoostAPR: 実行検証と二重報酬モデルによる自動プログラム修正の強化学習 BoostAPR: Boosting Automated Program Repair via Execution-Grounded Reinforcement Learning with Dual Reward Models

BoostAPR is a research effort aimed at improving large language model (LLM) based Automated Program Repair (APR) by combining reinforcement learning with execution-grounded feedback. The work targets a long-standing weakness in the field: producing patches that pass tests but fail to capture the true intent of the fix.

The approach rests on two main ideas. First, instead of training purely on textual similarity to reference patches, BoostAPR executes candidate patches against the project's test suite and feeds the outcome back as a learning signal. This execution-grounded reward is designed to discourage outputs that look syntactically plausible but do not actually compile or pass tests. Second, the system employs dual reward models. One reward estimates patch correctness, presumably leveraging test outcomes and fault-localization signals, while the other evaluates code quality dimensions such as minimality of edits and readability. Combining the two is intended to mitigate test overfitting, where a model learns to game a narrow test suite at the cost of semantic correctness.

The broader APR landscape provides useful context. Classical search-based tools like GenProg pioneered automated patching through genetic programming, while neural approaches such as CoCoNuT, SequenceR, and more recent LLM-driven systems like AlphaRepair and ChatRepair have steadily raised the bar on benchmarks such as Defects4J and QuixBugs. A persistent finding across these works is the gap between plausible patches (tests pass) and correct patches (intent preserved). BoostAPR's dual-reward structure can be read as a direct response to that gap, attempting to encode quality signals that pure pass/fail rewards cannot capture.

Reinforcement learning for code is itself an active area. Methods like CodeRL and execution-based RL for program synthesis have shown that grounding rewards in runtime behavior helps models generalize beyond surface patterns. More recent training recipes, including RLHF variants and group-relative policy optimization seen in systems like DeepSeek-Coder, point to a broader trend of treating code generation as a verifiable task where ground-truth execution can replace noisy human preferences. APR is arguably an even better fit for this paradigm than open-ended code synthesis, because bug-fix tasks come with concrete oracles in the form of failing tests.

Several caveats are worth noting. The quality of execution-grounded RL is bounded by the quality of the test suite; weak tests can still let incorrect patches through, and overly strict tests may suppress valid alternative fixes. Dual reward models also introduce the classic risk of reward hacking, where the policy exploits idiosyncrasies of the quality model rather than improving real code. As with most APR research, generalization across languages, project sizes, and bug categories remains an open question, and reported gains on curated benchmarks do not always translate to industrial codebases.

Because this summary is based on the arXiv listing, specific numbers, base models, and benchmark comparisons should be confirmed against the full paper. Still, the direction is consistent with where LLM-based program repair appears to be heading: less reliance on static supervised fine-tuning, and more on closed-loop training where the compiler and test runner act as the ultimate judge.