REA-Coder: A New Code Generation Method Bridging the Gap Between User Intent and Large Language Models via Requirement Alignment

This paper proposes the REA-Coder framework, whose core idea is to proactively identify and fix mismatches between requirements and large language models' understanding. By iteratively performing requirement alignment and code generation, it effectively addresses the intent gap problem in code generation. This method achieves an average improvement of 7.93% to 30.25% across five programming benchmark datasets, significantly outperforming existing baseline methods.

The core challenge in code generation lies in the ambiguity, incompleteness, or vagueness of users' natural language requirements, which leads to deviations between the model's understanding and the real intent (intent gap). Existing methods are divided into reasoning strategy optimization (e.g., chain-of-thought, tree search) and post-processing methods (e.g., test feedback repair), but all implicitly assume that the model can correctly understand the requirements, ignoring the premise of requirement understanding errors.

REA-Coder adopts an iterative loop of identification-alignment-generation-verification:

1. Mismatch Identification: Identify issues such as term ambiguity, implicit assumptions, and constraint conflicts in requirements through uncertainty quantification, adversarial probing, and knowledge graph matching;
2. Requirement Alignment: Explicitize implicit constraints, restructure in a structured way, supplement examples, and standardize terms to improve clarity while maintaining semantic equivalence;
3. Code Generation: Generate code based on aligned requirements;
4. Verification and Iteration: Perform verification of syntax, functionality, and requirement alignment; if it fails, return to iteration.

REA-Coder achieves performance improvements on four different large language models, proving its model-agnostic nature. It achieves an average improvement of 7.93% to 30.25% across five benchmark datasets, significantly outperforming baseline methods such as direct use of original requirements, simple prompt engineering, post-processing repair, and complex reasoning strategies.

Technical Insights: Requirement understanding is the bottleneck of code generation; iterative alignment embodies the idea of incremental refinement and implies a new mode of human-machine collaboration. Application scenarios include:

• Low-code/no-code platforms: Help non-professional users generate code that meets expectations;
• Intelligent programming assistants: Integrate into IDEs to reduce the number of modifications;
• Code generation API services: Improve service quality and user satisfaction.

Current Limitations: The quality of alignment relies on indirect evaluation via code verification. Future directions include: developing metrics for direct evaluation of alignment quality; extending to multi-turn dialogue scenarios; optimizing alignment strategies for specific programming domains; integrating explicit user feedback into the iterative loop.