MirrorCode Benchmark
Explore score history ↗25 target programs graded by exact output match against the reference on end-to-end tests, including held-out tests. Attempts are long-horizon and budget-sensitive: the paper reports $2,600 over 19 days for a single attempt on a large task, so two reported scores are comparable only when the inference budgets behind them are similar.
- Released
- 2026-06-29
- Evaluates
- coding agent
- Openness
- unknown
- Importer
- Model reports
MirrorCode paper
No reported scores are on record for this benchmark yet.
Independent audits of MirrorCode
- June Kim 2026-07-17
June Kim's audit reports that MirrorCode's claimed construct is autonomous reconstruction of whole programs from behavior, while its metric scores scoped reimplementation from a working oracle, partly via recall: the benchmark's own memorization screen flags 17 of 25 targets, and 2 of 25 are recall by construction (brotlid, mailauth). The audit discloses that Kim applied for a role at Epoch AI, which co-produced MirrorCode.