2026 · Preprint
MirrorCode: AI can rebuild entire programs from behavior alone
Tests reimplementation of 25 command-line programs using execute-only reference access, documentation and visible examples. Separate scoring environments compare exact outputs on visible and held-out end-to-end tests.
Scope and limitations
Targets are scoped, deterministic, stateless programs with text output. Test agreement is not universal equivalence or code-quality assessment. A memorization screen flags possible exposure for 17 targets. Large inference budgets and scaffold choices limit comparisons with cheaper benchmarks; three targets remain private.
Why it fits this collection
Directly evaluates sustained reconstruction from observable software behavior.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- June 29, 2026
- Date source
- First arXiv submission; the reviewed version is identified separately.
- First arXiv submission
- June 29, 2026
- arXiv
- 2606.30182
- Reviewed arXiv version
- v2
- DOI
- 10.48550/arxiv.2606.30182
June 29 first submission, July 17 v2: a six-month-window candidate. Distinct from ProgramBench, explicitly compared in the paper. The pinned repository has an MIT root license and public harness/task definitions, excluding private tasks; target-program licensing is separate. No benchmark was run.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.
Primary sources
- https://arxiv.org/abs/2606.30182
- https://arxiv.org/html/2606.30182v2
- https://api.datacite.org/dois/10.48550/arxiv.2606.30182
- README.md GitHub · epoch-research/MirrorCode · Revision 5c9d7b0
- LICENSE GitHub · epoch-research/MirrorCode · Revision 5c9d7b0