AI BriefingQwen Releases 'RecreationWorld' for Training AI Agents to Rebuild Apps
Key point
Qwen has released the RecreationWorld framework, which trains agents by using running apps as oracles.
Details
The Qwen team has released the RecreationWorld framework. This is a hybrid computer-use agent training system that trains AI agents to rebuild existing applications by iteratively performing GUI navigation, coding, and execution verification.
The core concept is to set the running reference app as an executable oracle, transforming open-source apps into scalable and verifiable learning experiences. Agents perform the explore, implement, and verify stages in an iterative loop rather than a fixed sequence.
The benchmark RecreationBench, released simultaneously, includes 250 test tasks and evaluates based on observable behavior rather than source code similarity. It provides 50 tasks each across five platforms: Linux (Ubuntu), macOS, Windows, Android, and Web.
In the benchmark results, GPT-6 Astra achieved the highest performance with an average of 58.06%, followed by Claude Opus 5 (44.16%) and GPT-5.6 Astra (42.06%). Qwen3.8-Max showed an average score of 34.80%.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Apple Releases LensVLM-9B for Processing Compressed Text Images
SWE-Bench Pro V2 Released: 89 Tasks Removed and Evaluation Protocol StrengthenedQwen Intelligence Launches Three Mobile AI Agents and Open Benchmark Suite