Enhancing queries for code generation with reinforcement learning

TL;DR

This paper introduces a reinforcement learning method to improve code generation by refining natural language queries, achieving significant performance gains.

Contribution

A novel reinforcement learning framework using LoRA to enhance code generation queries with combined text and execution rewards.

Findings

RL4QE improves code similarity by 34.3% on the DS1000 benchmark.

BLEU-4 is the most reliable text reward, and LoRA with rank 8 outperforms full fine-tuning.

Abstract

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 \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$r{=}8$$\end{document} 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.

Figures