Can We Translate Code Better with LLMs and Call Graph Analysis?
Authors: Yang Luo
IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on multiple mainstream datasets demonstrate that, compared to existing code translation methods and LLMs, this method achieves a significant improvement in translation accuracy. Conducting extensive experiments on multiple commonly used datasets, confirming that the Trans Graph method significantly improves the success rate of code translation compared to the existing baseline method, with an average increase of 15.7%. |
| Researcher Affiliation | Academia | Yang Luo1,2,3 1National Engineering Research Center for Software Engineering, Peking University, Beijing, China 2School of Software and Microelectronics, Peking University, Beijing, China 3PKU-OCTA Laboratory for Blockchain and Privacy Computing, Peking University, Beijing, China EMAIL |
| Pseudocode | Yes | Algorithm 1 Code Translation and Algorithm 2 Recursive Binary Search Debugging |
| Open Source Code | No | The paper does not contain any explicit statements about releasing code or links to code repositories for the methodology described. |
| Open Datasets | Yes | The datasets used in this paper include: Code Net [Puri et al., 2021], Avatar [Ahmad et al., 2021b], Eval Plus [Liu et al., 2024], Apache Commons CLI [apa, 2023] (Java), Click [pyt, 2023] (Python), and Human Eval-X [Zheng et al., 2023]. |
| Dataset Splits | No | The paper does not provide specific information regarding dataset splits (e.g., percentages, sample counts for training, validation, or testing sets). |
| Hardware Specification | No | The current Trans Graph implementation is primarily based on a single-machine multi-core CPU, but in scenarios with sufficient computing power, using GPUs or even multi-machine distributed environments can significantly improve the scale and efficiency of parallel translation. This sentence describes a general setup or future possibilities, not specific hardware used for the experiments. The paper does not provide specific CPU or GPU models or other detailed hardware specifications. |
| Software Dependencies | No | The compilation environments included gcc, Open JDK, and CPython, and the debugging tools included gdb, jdb, and pdb. Trans Graph integrates with VSCode using the Language Server Protocol to extract call graphs for these programming languages. In addition, it controls debuggers for various languages through the Debug Adapter Protocol. No specific version numbers are provided for these software dependencies. |
| Experiment Setup | Yes | The LLM prompts used in this experiment are as follows: Translate the following code from {source_language} to {target_language}. The translated code should be correct, efficient, and idiomatic. , {source_code} and To prevent prompts from becoming too long, we set a maximum length threshold (e.g., 32k tokens), and when this threshold is exceeded, low-priority contexts (such as comments) are truncated until the length limit is met. and Thirty experiments were conducted for each translation. |