Diagnosing Candidate Quality Bottlenecks in RL-Guided LLM Code Repair

Abstract

Recent Large Language Models (LLMs) have driven an intuitive expectation that automated program repair (APR) can be treated as a direct code-rewriting task. However, practical APR systems require generated candidates to be machine-usable, safely validated, and appropriately ranked. This paper uses an RL-guided repair loop as an instrumented diagnostic environment to separate correct-candidate availability from candidate selection. We evaluate all 40 Python programs in QuixBugs using four code-LLM families, four versioned output-contract conditions, and five fixed random seeds. The resulting 3,200 candidate-generation conditions were evaluated with public tests for plausible correctness and a separately defined 1,200-case augmented oracle for semantic correctness. All 40 programs and all prespecified runs were retained. Correct candidates were available in 173 of 3,200 conditions and for 30 of 40 programs, showing that availability was a necessary upstream constraint under the evaluated conditions. Static greedy selection improved the program-level correct-outcome rate over first selection by 10.0 percentage points (95% CI [0.0, 23.3]; Holm-adjusted $p=.25$ ), while the separate sequential comparison improved by 5.0 points (95% CI [0.0, 12.5]; Holm-adjusted $p=.50$ ); neither effect was statistically established. These conclusions are limited to single-file Python algorithmic repair on QuixBugs and do not establish general superiority of RL, a selector, or an output representation.