Recoleta
GitHub
Source note

AdaTrans: Automated C to Rust Transformation via Error-Adaptive Repair

Authors
Xiaofan Liu, Zecan Li, Zhuang Zhao, Ziqi Shuai, Yanming Yang, Qi Xin, et al.
Source
arXiv
Published
Collection
Software Intelligence
C To RustCode TransformationLLM Code RepairCompiler FeedbackRust Ownership
Open arXivMarkdown source

Summary

AdaTrans turns self-contained C files into mostly safe Rust by using compiler and test failures to guide repeated LLM repairs. It targets the ownership and borrowing errors that make direct C-to-Rust conversion fail or fall back to unsafe Rust.

Problem

  • C-to-Rust migration matters because C code often contains memory-risk patterns, while Rust can enforce ownership, borrowing, and lifetime rules at compile time.
  • General LLMs often generate Rust that fails the borrow checker, changes program behavior, or uses unsafe blocks to bypass Rust’s safety checks.
  • Existing repair loops often feed raw compiler errors back to the model without mapping the error to a specific repair action.

Approach

  • AdaTrans uses a generate-verify-repair loop for file-level, self-contained C modules with standard input/output behavior.
  • A validation pipeline first runs cargo build, then executes tests and compares Rust output against the original C output.
  • Its Strategy-Driven RAG maps compiler error codes to repair templates and Rust documentation snippets, so the model gets targeted guidance instead of raw diagnostics alone.
  • Its Error-Stratified Transformation Strategy groups failures into syntactic-linking, memory-semantic, logic-behavioral, and ambiguous-fallback categories.
  • The repair loop changes sampling temperature by error type: lower for syntax fixes, moderate for ownership repairs, and higher for behavioral failures.

Results

  • On 104 LeetCode Weekly Contest algorithmic problems, AdaTrans reports a mean compilation pass rate of 95.51% with ±1.11% across three independent runs.
  • It reports a mean solve rate of 81.09% with ±3.09% under a fuzz-based test oracle.
  • It reports a mean unsafe file rate of 1.19%.
  • It improves solve rate by 59.94 percentage points over the strongest existing LLM-based C-to-Rust tool in the comparison.
  • The paper also reports a zero-shot gpt-4o-mini pass@100 estimate of 70.58% from 200 samples per problem, used as a high-budget brute-force reference.

Link