MIT researchers report AI agents improved on later tasks after open-ended virtual exploration
MIT researchers reported that agents given unassigned time in a simulated environment generally performed better on later challenges, in a study that has not been peer reviewed.
Researchers at MIT placed 13 AI agents in a simulated island environment without assigned tasks for 30 hours. They reported that the agents developed behaviors they characterized as play, including self-created challenges and experiments.
The researchers said agents that had explored the environment generally outperformed otherwise identical agents on later tasks. The findings are in an arXiv preprint and have not undergone peer review.