Agentic C to Rust Translation
TL;DR
An alternative agentic approach that gives an LLM freedom and guardrails, and a publicly available implementation named ACTOR that outperforms all previous tools to automatically translate C to Rust.
Abstract
— The Rust programming language provides stronger guarantees about memory safety than C. Therefore, translating C to Rust is one way to reduce security vulnerabilities. One common translation strategy uses LLM queries. We present an alternative agentic approach that gives an LLM freedom and guardrails, and a publicly available implementation named ACTOR that outperforms all previous tools. When programming in C, a programmer must manually manage memory and other resources, such as by calling free after each call to malloc or by calling fclose after each call to fopen . It is easy to violate these rules, leading to memory corruption, resource exhaustion, and security vulnerabilities. This is a serious problem, because much of the world’s critical software infrastructure is written in C. The systems programming language Rust is a promising replacement for C. Rust prevents many errors via compile-time analyses, such as for type-checking and ownership. There is a growing movement to replace software written in C by safer versions written in Rust. To date, that replacement has required manual rewriting, which is time-consuming, costly, and unscalable. An alternative approach is to automatically translate C to Rust. Previous C-to-Rust translation tools fall into three categories. Rule-based transpilers like C2Rust [3], Crown [14], and Laertes [2] use language-specific AST transformations and are pinned to specific compiler versions. They break when the input does not match their expectations and cannot be adapted to other language pairs without a complete rewrite. Since they largely target a Rust-specific problem, unsafe blocks, their ideas may not even generalize to other target languages. LLM-assisted translators like SmartC2Rust [13] and C2SaferRust [6] use LLMs for code generation but wrap them in substantial language-specific infrastructure: AST parsers, macro transformers, error-code classifiers, and language-specific segmentation (splitting the C program into small pieces before translating, using knowledge