LLM4C2Rust: Large Language Models for Automated Memory-Safe Code Transpilation
New AI pipeline eliminates dangerous memory errors, converting legacy C code to secure Rust automatically.
A team of researchers has introduced LLM4C2Rust, a novel framework designed to automate the complex task of converting legacy C and C++ code into the memory-safe Rust programming language. The system employs a Retrieval-Augmented Generation (RAG) pipeline that strategically segments the input C/C++ code and augments a large language model (LLM) with relevant context retrieved from Rust documentation and compiler error references. This hybrid approach, which also integrates a smaller language model (SLM), guides the AI to produce more correct and secure transpilations than using an LLM alone.
The researchers tested their framework using several state-of-the-art OpenAI models, including GPT-4o, GPT-4-Turbo, and o3-Mini. The results demonstrated that the RAG-enhanced pipeline significantly improved both the functional correctness and security of the generated Rust code. Notably, when transpiling several Unix Coreutils programs, the framework achieved the complete elimination of critical memory-unsafe constructs like Raw Pointer Dereferences (RPDs) and Unsafe Type Casts (UTCs). This indicates a major step forward in using AI for automated software repair and modernization, potentially saving countless developer hours and reducing security vulnerabilities in critical legacy systems.
- Uses a RAG pipeline with models like GPT-4o to guide C-to-Rust transpilation using official documentation.
- Tested on Coreutils, it completely eliminated dangerous Raw Pointer Dereferences and Unsafe Type Casts in final Rust code.
- Proposes a hybrid LLM/SLM architecture with code segmentation to improve over rule-based or pure LLM approaches.
Why It Matters
Automates the secure modernization of billions of lines of vulnerable legacy C/C++ code, critical for infrastructure and embedded systems security.