MIT Researchers Find Play-Like Exploration Improves Claude Agents’ Performance on Virtual Island Tasks

MIT researchers found that Claude Opus 4.7 agents that spent 30 hours exploring a virtual island performed better on later tasks than fresh copies without that experience, though experts questioned whether their behavior qualified as play. The study has not yet been peer-reviewed. The agents, nicknamed Eko, invented games, arranged blocks into drawings, and set themselves challenges. Researchers later tested them on tasks such as getting as many objects as possible onto the island’s highest peak. Experienced agents outperformed fresh copies, and deleting memories reduced performance. Experts questioned the interpretation of the exploration as play, particularly because the prompts described Eko as curious and encouraged it to continue.

Generated by AI from reports · Updated 2 hours ago

There is not enough continuous observation data to plot a trend yet.