Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SRLCG: Self-Rectified Large-Scale Code Generation with Multidimensional Chain-of-Thought and Dynamic Backtracking

Hongru Ma, Yanjie Liang, Jiasheng Si, Weiyu Zhang, Hongjiao Guan, Chaoqun Zheng, Bing Xu, Wenpeng Lu

TL;DR

This paper introduces SRLCG, a framework for generating complete, multi-file software projects from a single prompt, addressing non-expert usability and the brittleness of prior code-generation methods. It combines Multidimensional CoT (strategic, tactical, operational) with a dynamic backtracking engine and adaptive self-rectification to ensure project-level coherence, correctness, and robustness. Across GPT-4 and DeepSeek-V3 backbones, SRLCG produces significantly longer, more complete projects and achieves notable gains in code length, completeness, correctness, usability, and robustness compared with strong CoT baselines. The work demonstrates practical potential for enabling non-programmers to generate robust, scalable software architectures, and points to future work on improving inference efficiency.

Abstract

Large language models (LLMs) have revolutionized code generation, significantly enhancing developer productivity. However, for a vast number of users with minimal coding knowledge, LLMs provide little support, as they primarily generate isolated code snippets rather than complete, large-scale project code. Without coding expertise, these users struggle to interpret, modify, and iteratively refine the outputs of LLMs, making it impossible to assemble a complete project. To address this issue, we propose Self-Rectified Large-Scale Code Generator (SRLCG), a framework that generates complete multi-file project code from a single prompt. SRLCG employs a novel multidimensional chain-of-thought (CoT) and self-rectification to guide LLMs in generating correct and robust code files, then integrates them into a complete and coherent project using our proposed dynamic backtracking algorithm. Experimental results show that SRLCG generates code 15x longer than DeepSeek-V3, 16x longer than GPT-4, and at least 10x longer than other leading CoT-based baselines. Furthermore, they confirm its improved correctness, robustness, and performance compared to baselines in large-scale code generation.

SRLCG: Self-Rectified Large-Scale Code Generation with Multidimensional Chain-of-Thought and Dynamic Backtracking

TL;DR

This paper introduces SRLCG, a framework for generating complete, multi-file software projects from a single prompt, addressing non-expert usability and the brittleness of prior code-generation methods. It combines Multidimensional CoT (strategic, tactical, operational) with a dynamic backtracking engine and adaptive self-rectification to ensure project-level coherence, correctness, and robustness. Across GPT-4 and DeepSeek-V3 backbones, SRLCG produces significantly longer, more complete projects and achieves notable gains in code length, completeness, correctness, usability, and robustness compared with strong CoT baselines. The work demonstrates practical potential for enabling non-programmers to generate robust, scalable software architectures, and points to future work on improving inference efficiency.

Abstract

Large language models (LLMs) have revolutionized code generation, significantly enhancing developer productivity. However, for a vast number of users with minimal coding knowledge, LLMs provide little support, as they primarily generate isolated code snippets rather than complete, large-scale project code. Without coding expertise, these users struggle to interpret, modify, and iteratively refine the outputs of LLMs, making it impossible to assemble a complete project. To address this issue, we propose Self-Rectified Large-Scale Code Generator (SRLCG), a framework that generates complete multi-file project code from a single prompt. SRLCG employs a novel multidimensional chain-of-thought (CoT) and self-rectification to guide LLMs in generating correct and robust code files, then integrates them into a complete and coherent project using our proposed dynamic backtracking algorithm. Experimental results show that SRLCG generates code 15x longer than DeepSeek-V3, 16x longer than GPT-4, and at least 10x longer than other leading CoT-based baselines. Furthermore, they confirm its improved correctness, robustness, and performance compared to baselines in large-scale code generation.

Paper Structure

This paper contains 41 sections, 6 equations, 3 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. CoT-Based Code Generation
  4. Self-Checking in Code Generation
  5. Methodology
  6. Task Definition
  7. Overview
  8. MdCoT-DB
  9. Multidimensional CoT
  10. Dynamic Backtracking Algorithm
  11. Self-Rectification
  12. Experiments
  13. Experimental Setup
  14. Performance Comparison with Baselines
  15. Code Length Evaluation
  16. ...and 26 more sections

Figures (3)

  • Figure 1: Comparison of SRLCG and existing methods. Existing methods struggle with multi-file project generation while ensuring correctness and robustness, whereas SRLCG directly generates correct and robust large-scale project code.
  • Figure 2: The SRLCG framework consists of two modules: (a) MdCoT-DB, which decomposes user prompts into sub-functions and then integrates them via dynamic backtracking, and (b) Self-Rectification applies to each MdCoT dimension, performing verification and rectification to ensure the correctness and robustness of the generated code.
  • Figure 3: Analysis of parameters in the SRLCG framework.