LLM Feedback Loops and Code Perturbations in C-to-Rust Translation
A recent arXiv paper (2512.02567v2) examines the impact of automated feedback loops, the choice of Large Language Model (LLM), and behavior-preserving code changes on an automated C-to-Rust translation system. The system employs a generate-and-check pattern, where LLM-generated Rust code is automatically verified for compilability and behavioral equivalence. The study is motivated by Rust's safety guarantees, making C-to-Rust translation attractive for industry. The research focuses on three variables: feedback loops, LLM selection, and code perturbations, and their influence on translation quality. The paper is part of the ongoing exploration of generative AI in software engineering, particularly for code repair, test generation, and language translation. While interactive tools like GitHub Copilot are widely used, automated approaches require higher reliability for industrial adoption. The findings aim to inform the development of more robust automated translation systems.
Key facts
- arXiv:2512.02567v2 is a paper on LLM-based software engineering.
- The study focuses on a C-to-Rust translation system.
- Three variables are studied: automated feedback loops, LLM choice, and behavior-preserving code changes.
- The translation system uses a generate-and-check pattern.
- Rust's safety guarantees make C-to-Rust translation attractive for industry.
- The paper is motivated by the need for higher reliability in automated approaches.
- GitHub Copilot is mentioned as an interactive tool already in widespread use.
- The paper was announced as replace-cross on arXiv.
Entities
Institutions
- arXiv