FLARE: Fine-Grained Diagnostic Feedback for LLM Code Refinement

摘要

Large language models often generate code with bugs. Existing methods revise such code using feedback signals such as test failures and self-critiques, but these signals are too coarse-grained or too high-level to tell the model where the code should be revised. We present Flare, an iterative code generation framework whose lightweight diagnostic model predicts line-level suspiciousness to guide code refinement. Given the inherent uncertainty of diagnostic predictions, Flare searches over the top-k suspicious lines at each step and selects the next candidate according to execution outcomes. Experiments on LiveCodeBench and BigCodeBench with five base LLMs show that, even without candidate search, Flare ( $k$ =1) outperforms the strongest baseline, NL-Debugging, with an absolute improvement from 1.06% to 9.71% in 9 of 10 settings. Expanding the search to 10 candidates adds an average improvement of 8.97%, and Flare keeps an average lead of 3.39% over the strongest baseline under the same candidate budget. When evaluated in isolation, our lightweight diagnostic model outperforms recent fault localization methods, demonstrating that it can provide reliable fine-grained guidance for code refinement.