Enhancing queries for code generation with reinforcement learning

• Published in: Scientific reports Vol. 15; no. 1; pp. 37300 - 12
• Main Authors: Yuan, Dawei, Liang, Guojun, Li, Tingting, Liu, Suping
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 24.10.2025; Nature Publishing Group; Nature Portfolio
• Subjects: 639/166; 639/705; Code generation; Cognition & reasoning; Efficiency; Entropy; Feedback; Humanities and Social Sciences; Language; Large language models; Methods; multidisciplinary; Natural language; Optimization; Parameter-efficient fine-tuning; Prompt engineering; Queries; Reinforcement; Reinforcement learning; Science; Science (multidisciplinary); Code generation; Parameter-efficient fine-tuning; Prompt engineering; Reinforcement learning
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-21271-4

We present a reinforcement learning framework that enhances natural language queries to improve DeepSeek code generation. A parametric refiner (Qwen with LoRA) is trained via REINFORCE while the generator remains fixed, using a scalar reward that can combine text similarity (BLEU-4, ROUGE-L, F1, Overlap) with execution signals (unit tests, syntax/timeout penalties). On the DS1000 benchmark (800 train / 200 test), RL4QE improves the code similarity by 34.3%. Ablations show that BLEU-4 is the most reliable text reward overall (with F1 competitive on a larger scale), and LoRA with rank outperforms complete fine-tuning on most metrics while being more parameter efficient. The approach is transferred across foundation models (e.g., Qwen1.5/2/2.5 variants), where architecture often matters more than size. RL4QE is easy to integrate in practice (LoRA in attention projections) and supports reproducibility.