DeepEvolve: self-improving forecaster tops FutureX for 3 weeks with up to 195% rating gain
Zhipeng Chen · hf · 2026-10-22
Researchers from Renmin University's Gaoling School of AI, IQuest Research, and NTU introduce DeepEvolve, an agent system for recursive self-improvement (RSI) grounded in future event prediction.
- Motivation: Most AI systems ship with frozen parameters and improve only via external updates; studying RSI requires an environment that is continuously changing, objectively verifiable, and naturally grounded.
- Design: A living environment synthesizes verifiable tasks, while a recursive-evolution engine converts each prediction's reasoning trajectory back into capability — decoupling search from prediction to resist contamination, and distilling reusable skills to compensate for scarce live trajectories.
- Results: On the weekly-updated FutureX leaderboard scoring predictions against real outcomes, the self-improved forecaster ranked first for three consecutive weeks with up to 195% rating improvement; on synthesized evals it placed top-3 across four base LLMs, beating the strongest baseline by up to 10.7 Bradley-Terry points.
The authors frame this as a shift from training static models on fixed data toward agents that grow stronger from their own experience in open worlds.
More from AGI Musings
- You don't need superintelligence to break the digital job market — human-level AGI is enough — flowersslop · 2026-10-22
- Crypto hacks are the canary: attack counts surge but dollar losses stay flat — markjeffrey · 2026-10-22
- "We are here": cocktailpeanut posts Forster's The Machine Stops — cocktailpeanut · 2026-10-22
- 'Machines are doing original math now' — the countdown for desk-thinking jobs — rand_longevity · 2026-10-22
- 'If we lose the internet forever tonight': researchers joke—and stress—about living through the singularity — ZeroStateReflex · 2026-10-22
- If attackers and defenders both have ASI, what's the equilibrium? — flowersslop · 2026-10-22