PaperCompiler: Faithful Paper-to-Code Generation via Repository-Level Specification Compilation

PaperCompiler. Faithful Paper-to-Code Generation via Repository-Level Specification Compilation. By Yunhao Liu and two others, from Nanyang Technological University, Singapore. Published on the second of September. Its code repository is public and real: eleven source files.

Turning a research paper into a working repository is hard to do faithfully. Papers often describe methods at a high level and leave implementation assumptions implicit, while the repository that comes out has to preserve the method's logic, its evaluation protocol, and consistency across many files. Paper-to-code agents have advanced, but their intermediate outputs tend to be free-form plans or summaries. A downstream coding agent may ignore those, reinterpret them, or compress them, and the result is a simplified algorithm and a repository whose structure does not hang together.

PaperCompiler replaces the loose plan with something closer to a contract. It compiles evidence grounded in the paper into explicit, repository-level implementation specifications. While doing so it keeps track of where each piece of evidence came from, and it separates four kinds of information: what the paper supports, what was inferred, what is delegated to something external, and what remains unresolved.

The compiled specifications encode several things a free-form plan leaves loose. There are non-degradation requirements, ownership assignments, dependencies that cross file boundaries, and constraints at the level of individual files. Repository generation then proceeds under those specifications. It is not a rigid template, though. Local engineering choices that the paper itself does not fix are left flexible, so the specification binds what the paper determines and leaves the rest to the coding agent.

On Paper2CodeBench, PaperCompiler outperforms strong baselines. Reference-based fidelity rises from three point six four to four point one five, a relative improvement of thirteen point eight percent. High-severity critiques from the evaluator fall from thirteen point two percent to six point one percent, less than half the earlier rate.

The contribution sits in the hand-off between reading and writing. Where earlier pipelines passed a summary that could be quietly reinterpreted, this one passes a specification that records provenance, marks what is known and what is not, and states which requirements must survive into the code.