Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code

The paper introduces generative compilation, a method for obtaining compiler feedback while an autoregressive model is still generating code. Its central device, called a sealor, transforms a partial program into a complete program that standard compilers can diagnose. The construction is designed never to reject a partial program that could still be completed, while retaining enough context to identify genuine dead ends early. The authors prove these properties for a core Rust-like calculus with mechanized Lean verification, extend the approach to real Rust, and evaluate repository-level tasks using both frontier black-box and open-weight models. The reported results show fewer non-compiling outputs and better functional correctness than standard post-generation feedback.