Decaf: Compiler Feedback Improves Neural Decompilation via Search

A new system called Decaf (DECompilation with Automated Feedback) uses compiler feedback to improve the semantic correctness of neural decompiler outputs. Decompilers reverse-engineer compiled binaries back into source code, a task complicated by the loss of high-level syntax, identifiers, and custom data types during compilation. While deterministic decompilers are useful, they struggle with idiomatic syntax and identifier names. Generative AI models can reconstruct high-level constructs but often hallucinate improper programming constructs and semantics. Instead of relying on more data or training, Decaf employs a search process guided by compiler feedback to enhance output correctness. The approach is detailed in a paper on arXiv (2605.11501).