DeepEvolve agents self-improve by forecasting live events
Original title: Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting
Authors say the forecaster led FutureX for three weeks, with rating gains up to 195%.
The paper presents DeepEvolve, an evolution system in which agents predict future events, interact with a changing verifiable environment, and improve by internalizing and externalizing that experience. A living environment synthesizes tasks, while a recursive-evolution engine turns prediction trajectories into capability by separating search from prediction and distilling reusable skills. On the weekly FutureX leaderboard, the authors say their self-improved forecaster ranked first across all systems for three consecutive weeks, with up to a 195% rating improvement. On a synthesized evaluation it ranked in the top three across four base LLMs.
Unconfirmed: Leaderboard rank and the 195% rating gain are the authors' claims, not independently confirmed here.
- Search is decoupled from prediction to limit live-data contamination.
- Skill distillation lets the agent learn from limited live trajectories.
- Synthesized eval beat the strongest baseline by up to 10.7 Bradley-Terry points.
- Results cover four base LLMs on the authors' synthesized evaluation.
From the original article
1 min read at huggingface.co
Most contemporary artificial intelligence (AI) systems are released with frozen parameters and improve through external updates, whereas an ideal system would keep improving from its own experience, a capacity we call recursive self-improvement (RSI).
By Zhipeng Chen, Yunsen Liang, Yanxi Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.