ProgramBench tests cleanroom software reconstruction

Clever benchmark, dangerous incentive. It measures whether an agent can infer a large program’s behavior from black-box evidence, but it also invites teams to optimize for the eval rather than for useful coding skill.

• –The setup is intentionally hard: execute-only binaries, hidden behavioral tests, and no internet make shortcutting much harder than on ordinary coding benches.
• –That makes ProgramBench more representative of systems work than toy patch tasks, but also more sensitive to scaffolding, search strategy, and test-time agent design.
• –Early public results look brutal, with no fully solved tasks and only tiny “almost resolved” rates at the top of the leaderboard.
• –If people start overfitting to ProgramBench, the benchmark will reward benchmark-specific reverse-engineering tricks, not better software engineers.