D2Coder: large language models based agent for coding with dynamic debugging tools
Keywords: LLM-based Agent, Call Graph, Dynamic Debugging, Fault Localization
Abstract: Intelligent agents based on large language models have demonstrated certain programming abilities, but there is still significant room for improvement in complex project-level debugging tasks. Previous work has utilized general multi-agent workflows to enhance performance but has the following issues: 1) excessive reliance on the reasoning capabilities of large language models without debugging and detailed analysis of the code; 2) lack of intrinsic code information, such as call relationships and dependencies; 3) insufficient analysis and optimization of critical stages, especially the code search capability in fault localization, which directly affects the effectiveness of subsequent stages. Based on the SWE-bench dataset, we first isolate the fault localization capability for separate analysis and experiments, and introduce program call graphs to demonstrate the effectiveness of this information for debugging. Furthermore, during the debugging phase, we propose a simulated debugging mode that enables large language models to simulate program debugging without relying on other debugging tools. Compared to the real machine debugging mode, our experiments prove the effectiveness and generality of the simulated debugging mode. We conducted experiments on SWE-bench and improved the resolution rate by approximately 27.3\%, demonstrating the potential of this method.
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 6312
Loading