Evolving with Open Worlds: Towards Recursive Self-Improvement for Future Forecasting
The authors report that DeepEvolve led the FutureX forecasting leaderboard for three consecutive weeks, with rating gains of up to 195%.
DeepEvolve uses future-event prediction as an environment where agents can learn from outcomes that can be checked. It feeds prediction reasoning back into model updates and distills reusable skills from limited live experience. In the authors’ synthesized evaluation across four base language models, it placed in the top three and beat the strongest baseline by up to 10.7 Bradley–Terry points. The paper is marked as work in progress.
HF Daily Papers' note
DeepEvolve uses future-event prediction as an environment where agents can learn from outcomes that can be checked. It feeds prediction reasoning back into model updates and distills reusable skills from limited live experience. In the authors’ synthesized evaluation across four base language models, it placed in the top three and beat the strongest baseline by up to 10.7 Bradley–Terry points. The paper is marked as work in progress.
HF Daily Papers' note
score 4