You can deploy this model and start earning money today!

docz1105/
MultiFork

C++/C Code

Compiler Backend Development

code

cc-by-4.0

Views

No views yet

A Multi-Modal Retrieval-Augmented Framework for Compiler Backend Generation with LLMs

This project provides the scripts, datasets and fine-tuned models for reproducing the results in the paper titled "A Multi-Modal Retrieval-Augmented Framework for Compiler Backend Generation with LLMs".

MultiFork is a multi-modal representation learning RAG framework for enhancing LLMs compiler backend generation accuracy.

1. Dependency

Hardware
- A 64-core Intel Xeon Gold CPU
- A Nvidia A100 GPU (80GB Memory)
Software There are two conda virtual environments in our experiments, run the following commands to create both environments.

shell
1conda env create -f ./SimGNN.yml 2conda env create -f ./unixcoder.yml

2. Reproducing the Results

2.1 Table.1

Run following command for generating results for Table.1:

bash ./Script/LLMs/run_LLMs.sh

The results will be saved in ./Script/LLMs/Res_AE/*.csv.

2.2 Table.2

Run following command for generating results for Table.2:

bash ./Script/CodeT5p/run_codet5p.sh

The results will be saved in ./Script/CodeT5p/Res_AE/codet5p_func.csv.

2.3 Fig.7

Run following command for generating results for Fig.7:

bash ./Script/VEGA_Enhanced/run_vega.sh

The results will be saved in ./Script/VEGA_Enhanced/Res_AE/LLMs_Result.csv.

2.4 Fig.8

Run following command for generating results for Fig.8:

bash ./Script/Target-Level/run_target_level.sh

The results will be saved in ./Script/Target-Level/Res_AE/target_level.csv.

2.5 Table.3

Run following command for generating results for Table.3:

bash ./Script/Func-Level/run_func_level.sh

The results will be saved in ./Script/Func-Level/Res_AE/func_level.csv.

2.6 Fig.9

Run following command for generating results for Fig.9:

bash ./Script/Aba_Study/Target-Level/run_target_aba.sh

The results will be saved in ./Script/Aba_Study/Target-Level/Res_AE/target_aba.csv.

2.7 Table.4

Run following command for generating results for Table.4:

bash ./Script/Aba_Study/Func-Level/run_func_aba.sh

The results will be saved in ./Script/Aba_Study/Func-Level/Res_AE/func_aba.csv.

Citation

@inproceedings{zhong2026multifork,
  title={A Multi-Modal Retrieval-Augmented Framework for Generating Compiler Backends with LLMs},
  author={Ming Zhong, Fang Lv, Hongna Geng, Xin Sun, Lulin Wang, Lei Qiu,  Huimin Cui, Xiaobing Feng},
  booktitle={2026 IEEE 33rd International Conference on Software Analysis, Evolution and Reengineering (SANER)},
  year={2026}
}

docz1105/MultiFork