FutureWorld: A live reinforcement learning environment for training AI agents on real-world prediction tasks

Researchers introduced FutureWorld, a live agentic reinforcement learning environment designed for live future prediction—the task of making predictions about real-world events before they unfold. The environment closes the training loop between prediction, outcome realization, and parameter updates.

The system trains three open-source base models over consecutive days and provides a unified learning environment that grounds prediction questions in diverse real-world events while preventing answer leakage. The authors built a daily benchmark and evaluated several frontier agents to establish performance baselines.

Training in the environment was effective, according to the results. The work frames live future prediction as a learning environment for building agents that can continually learn from real-world outcomes.