← home

A Study on Robustness and Reliability of Large Language Model Code Generation

2024/01/14

Reference: https://arxiv.org/pdf/2308.10335.pdf

In short:Currently, software engineers should use tools like Co-pilot cautiously.

The evaluation results show that even for GPT-4, 62% of the generated code contains API misuses, which would cause unexpected consequences if the code is introduced into real-world software.

Therefore, it is essential to contemplate the code reliability while evaluating the code generation by large language models.

Result of checking API Usage from LLMs - Figure 3

For LLM Practitioners:

  • Improvement in the base model helps.
  • Fine-tuning the model helps.
  • Prompt engineering helps.

Findings:

  1. Answers to real-world coding questions from state-of-the-art large language models often have API misuse issues.
  2. Among all answers containing executable code, 57-70% of the code snippets in LLM answers contain API misuse, potentially leading to severe consequences in production.
  3. Irrelevant short examples do not decrease the API misuse rate but trigger more valid answers, proving effective for benchmarking model performance.
  4. Some LLMs can learn from correct usage examples, reducing the API misuse rate.
  5. GPT-4 produces the highest number of answers containing executable code. Different LLMs show varying misuse rates for benchmark APIs.

#llm