20 Top Global Universities Jointly Launch ScienceIDE to Train AI with Real Scientific Outcomes

Researchers from 20 top global universities, including Oxford, Stanford, and Princeton, have jointly launched ScienceIDE, a project that organizes scientific code, runtime environments, and validation rules into tasks for AI models to execute, verify, and learn from. The project found that among 1,769 defect tasks, 55.2% of programs compiled and ran successfully but produced incorrect scientific results, highlighting the need to constrain AI behavior with real scientific outcomes to prevent models from treating errors as successful rewards.