Improve Code Generation with Feedback

Zhi Xu, Yun Fu

Keywords: LLM, code generation

Abstract:

As advancements in Large Language Models (LLMs) continue to accelerate, an increasing number of researchers are exploring the potential of these models to assist in everyday tasks. Despite their remarkable achievements in various downstream applications, several challenges must be addressed. This paper delves into applying LLMs in coding tasks, such as ChatGPT and LLama. Initial observations suggest that directly employing these LLMs does not yield optimal results. However, we have identified that LLMs demonstrate enhanced performance when given appropriate feedback. This includes providing information on the accuracy of the code generated, supplying test cases relevant to the task, and indicating the correct or incorrect outputs for these test cases.
Furthermore, we have developed an innovative architecture miming human debugging. This approach supplies local variable information to the LLM while executing the generated code. Our architecture facilitates providing feedback to the LLM and simulates the human debugging experience, thereby significantly improving the LLM's code generation capabilities.
Utilizing our proposed architecture, our model surpasses the current benchmarks of state-of-the-art models in the MBPP and Humaneval datasets. We also present comprehensive analyses and ablation studies to substantiate the efficacy of our methods. These findings open new avenues for enhancing the utility of LLMs in coding tasks, offering a more interactive and practical approach to leveraging these advanced technologies.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10967