Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting

Evolving with Open Worlds. Towards Recursive Self-Improvement for Future Forecasting. By Zhipeng Chen and fourteen others, from Renmin University of China, IQuest Research, and one other institution. Published on the twenty-first of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Most artificial intelligence systems today are released with frozen parameters. They get better only when someone updates them from outside. The authors describe a different ideal: a system that keeps improving from its own experience, a capacity they call recursive self-improvement, or RSI. The obstacle to studying it is the environment. To learn from experience indefinitely, a system needs a world that keeps changing, that can be checked objectively, and that is naturally grounded rather than contrived. Finding an environment with all three properties at once is described as a fundamental challenge for agent evolution.

Future event prediction is the answer the authors choose. Tomorrow's events are always new, and once they happen they can be verified against what actually occurred. On that footing they build DeepEvolve, a sustainable evolution agent system. Its agents interact continuously with a dynamically evolving environment, and they improve by internalizing and externalizing the experience those interactions produce. Improvement is not a separate training phase applied from outside. It comes out of the agent's own record of making predictions and seeing how they turned out.

Two parts carry the design. A living environment synthesizes verifiable tasks, so there is always fresh material to predict. A recursive-evolution engine then takes the reasoning trajectory behind each prediction and turns it back into capability. Two choices inside that engine answer two difficulties of a live setting. The first is contamination. A live environment is prone to it, so the engine decouples search from prediction, which is what lets parameter-level evolution work there at all. The second is scarcity. A live environment affords only a limited number of trajectories, so the engine distills reusable skills from them, letting the agent evolve without needing vast amounts of experience.

The external test is the FutureX leaderboard, which scores predictions against real-world outcomes and is updated every week. The self-improved forecaster ranked first across all systems there for three consecutive weeks, with up to a one hundred and ninety-five percent improvement in rating.

On the authors' own synthesized evaluation, it ranked among the top three across four different base LLMs, and it outperformed the strongest baseline by up to ten point seven Bradley Terry points.

The authors draw a broader reading from these results. A system built for self-improvement, they argue, is best understood as an agent that co-evolves with a genuinely living, open-ended environment. That points to a shift away from training static models on fixed data, toward cultivating agents that grow stronger from their own experience. In their account the open world is not a convenient testbed for that growth. It is the indispensable basis of it, which is why the environment, and not the model, sits at the centre of the whole design.