The paper presents Flare, a lightweight diagnostic model for predicting line-level suspiciousness signals to help large language models refine generated code. Experiments show that Flare outperforms baseline methods, achieving an average improvement of 8.50% with candidate search.