Benchmark
RecreationBench
Test whether an agent can inspect a working application and build a replacement that behaves and looks like it.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The RecreationWorld team assembled 250 held-out tasks, with 50 each for Ubuntu, macOS, Windows, Android and the web. The linked report credits Shuai Bai and 32 coauthors.
Methodology
The agent explores a running reference app and writes its own implementation. The evaluator tests visible behavior and visual fidelity. Reference source code and evaluation tests are hidden from the agent. RecreationWorld provides the execution framework. RecreationBench is its held-out suite. The Qwen model card reports use of the suite without identifying a numbered benchmark edition.