rsi.watch

Zhipeng Chen et al. (15 authors) · 2026-10-21

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

“our self-improved forecaster ranked first across all systems over three consecutive weeks, achieving up to a 195% improvement in rating”
Who
Zhipeng Chen et al. (15 authors)
Date
2026-10-21
Evidence
Moderate certainty Starts at High: measured by someone other than the developer.
  • ↓ indirect measure. FutureX ranks the whole forecasting system against other entrants. The paper calls its reported results a system "after the full recursive self-improvement", and calls the FutureX frontier "where our agent begins rather than ends". “This frontier is where our agent begins rather than ends: the remaining subsections show that recursively evolving Branchus pushes it beyond this already-leading base form”
how we rate evidence
Replication
No rerun found. The FutureX ratings are computed by the benchmark. We found no public code for the system and no rerun by another group. The ratings are read from the authors' captures of the FutureX site, where the entry is listed as self-submitted. Searched up to 2026-10-26. all claims
Checked
2026-10-26, against the source page
Links
Source Paper

Also on the page

  • The paper reproduces four FutureX leaderboards it calls weekly, dated 2026-08-17, 2026-08-26, 2026-09-03 and 2026-09-07. The system was first in the first three; in the last it "remains among the strongest entrants". In the authors' 2026-09-07 snapshot it is second, rated 0.83 against 0.99.
  • Rank is by BT rating. In the 2026-08-17 snapshot the entry is rated 1.15 and the next 0.39. The paper's text does not say what the 195% is measured against.
  • The paper does not name the base model used on FutureX. In the authors' snapshots the entry is Branchus-beta on GLM-5.1. FutureX entrants "run your model locally, and email us your predictions" each week.
  • Version 1 called the system DeepEvolve and said "Our code and data are available" on GitHub. The authors' EvoForesight repository holds a source list and leaderboard snapshots; its Quick Start section is empty.

← Ledger Report an error