Search papers, labs, and topics across Lattice.
This paper introduces generative compilation, a novel approach that provides real-time compiler feedback during the code generation process of AI models, specifically targeting the challenges posed by languages with strict static semantics like Rust. By employing a lightweight transformation called a sealor, the method ensures that partial programs can be completed and diagnosed by standard compilers without rejecting potentially valid tokens. The results demonstrate that generative compilation significantly reduces non-compiling outputs and enhances functional correctness, outperforming traditional post-generation feedback mechanisms.
Generative compilation enables AI models to receive compiler feedback during code generation, drastically reducing errors and improving code quality in real-time.
Languages with rich static semantics, such as Rust, provide stronger guarantees for AI-generated code, but their strictness makes generation more difficult. Off-the-shelf compilers can provide useful feedback post-generation, but does not guide intermediate generation steps, such as those during autoregressive LLM decoding. Constrained decoding intervenes earlier by rejecting invalid tokens during sampling, but requires white-box model access and costly reimplementation for semantic constraints. We introduce generative compilation, the first approach to obtaining compiler feedback on partial programs during generation. The core technical device is a sealor: a lightweight, mostly syntax-guided transformation that converts partial programs into complete ones that standard compilers can diagnose. It is designed such that possible-to-complete partial programs are never rejected, while preserving enough code context to catch genuine dead ends early. We construct such a sealor on a core Rust-like calculus and prove that it satisfies these properties, all mechanized in Lean. We extend it to the first partial-program checker for real Rust. We evaluate our method on challenging repository-level Rust coding tasks, across both frontier black-box and open-weight models. We show that generative compilation reduces non-compiling outputs and improves functional correctness, relative to standard post-generation feedback. It does so by detecting a broad range of errors close to their source and early during generation, thereby reducing errors cascades and enabling focused diagnostics. More broadly, generative compilation is a step toward making compilers a first-class citizen of AI-assisted programming active during generation, rather than a separate post-generation check.
