[Paper] Bridging the Gap between User Intent and LLM: A Requirement Alignment Approach for Code Generation

Source: arXiv - 2604.16198v1

Overview

The paper introduces REA‑Coder, a novel “requirement alignment” loop that bridges the gap between what a user asks for and what a large language model (LLM) actually understands when generating code. By iteratively detecting mismatches, clarifying the intent, and re‑prompting the model, REA‑Coder consistently produces more correct programs across several popular code‑generation benchmarks.

Key Contributions

• Requirement‑Alignment Loop: A systematic process that first spots mis‑understood parts of a user’s specification, then refines the prompt before code generation.
• Iterative Verification: After each generation, the produced code is checked against the aligned requirements; mismatches trigger another alignment round.
• Model‑agnostic Design: Works with any LLM capable of code generation (tested on four models) without requiring model‑specific fine‑tuning.
• Empirical Gains: Achieves average improvements of 7.9%–30.3% over strong baselines on five widely used programming benchmarks.
• Open‑source Toolkit: The authors release the REA‑Coder pipeline, making it easy for developers to plug into existing LLM‑based coding assistants.

Methodology

1. Initial Prompt & Generation – The user’s natural‑language requirement is fed to an LLM, which returns a candidate program.
2. Alignment Check – A lightweight verifier (either a rule‑based parser or a secondary LLM) compares the generated code against the original requirement, flagging any mismatched intent (e.g., missing edge‑case handling, wrong API usage).
3. Requirement Refinement – The system automatically rewrites the ambiguous parts of the prompt, adding clarifying details or constraints identified in step 2.
4. Re‑generation – The refined prompt is sent back to the LLM.
5. Iterative Loop – Steps 2‑4 repeat until either the code passes all alignment checks or a preset iteration limit is reached.

The loop is lightweight: verification uses static analysis or unit‑test execution, and prompt refinement leverages the same LLM, keeping the pipeline fully automated.

Results & Findings

• Consistent gains across four LLMs (GPT‑3.5, Claude‑2, LLaMA‑2‑13B, CodeGen‑6B).
• Fewer failed test cases: the verification step eliminates many “off‑by‑one” or missing‑requirement bugs that standard generation pipelines overlook.
• Iteration budget: Most successful runs converge within 2–3 iterations, keeping latency acceptable for interactive use.

Practical Implications

• Better Coding Assistants: Integrating REA‑Coder can turn a “write‑me‑a‑function” chat into a more reliable tool that self‑corrects before presenting code to the developer.
• Reduced Manual Debugging: By catching requirement mismatches early, developers spend less time hunting down logical errors that stem from misunderstood specs.
• Safer Automation: In CI pipelines that auto‑generate scripts (e.g., data‑pipeline scaffolding), the alignment loop acts as a guardrail, ensuring generated code truly satisfies the declared contract.
• Low‑cost Adaptation: Since REA‑Coder works with off‑the‑shelf LLMs, teams can boost existing services without expensive fine‑tuning or model retraining.

Limitations & Future Work

• Verification Dependency: The quality of the alignment check hinges on the availability of reliable test cases or static analysis rules; for highly ambiguous or novel domains, this may be weak.
• Latency Overhead: Each iteration adds an extra LLM call; while most cases converge quickly, latency‑sensitive applications may need tighter iteration caps.
• Prompt Drift: Re‑writing prompts automatically can occasionally over‑specify, limiting the model’s creative solutions.
• Future Directions: The authors suggest exploring learned verification models, adaptive iteration budgets, and extending the loop to multi‑module codebases or UI‑driven specifications.

Authors

• Jia Li
• Ruiqi Bai
• Yangkang Luo
• Yiran Zhang
• Wentao Yang
• Zeyu Sun
• Tiankuo Zhao
• Dongming Jin
• Lei Li
• Zhi Jin

Paper Information

• arXiv ID: 2604.16198v1
• Categories: cs.SE
• Published: April 17, 2026
• PDF: Download PDF