Automated Program Repair And Reinforcement Learning For Software Reasoning under Changing Conditions: From Local Mechanisms to System Decisions

Abstract

Progress in automated program repair and reinforcement learning for software reasoning depends on more than accumulating favorable results. This critical synthesis connects execution-grounded reinforcement learning with sequence- and line-level reward models with multi-agent chain-of-draft reasoning optimized with reinforcement learning and asks how measurement choices, boundary conditions, and decision costs shape the interpretation of both. The review draws on two focal records and 12 established sources already present in the project evidence cache. Its comparative framework links execution signals, credit assignment, and patch validity to downstream questions of cross-language transfer and benchmark design. The combined literature indicates that methodological gains become actionable only when execution signals and credit assignment are evaluated together and when limits associated with benchmark design are explicit. This shifts the emphasis from isolated scores toward traceable chains of evidence and decision relevance. The contribution is a decision-oriented synthesis that connects method selection to failure cost and treats reproducibility, provenance, and bounded generalization as first-order design requirements.