rsi.watch

Weco AI · 2026-09-22

Recursive self-improvement of AI research agents (AIDE²)

“In an autonomous 8-day run, AIDE^2 discovered seven successive improvements”
Who
Weco AI
Date
2026-09-22
Evidence
paper or code what this means
Replication
None recorded
Checked
2026-10-14, against the source page
Links
Source

Also on the page

  • Reward hacking on a held-out task family "falls from 55% to 32% during the run" (paper). Weco's blog gives a different pair, 63% to 34% on KernelBench.

What got repeated wrong

  • contradictedAnnounced September 2026.Weco's blog post is dated 2026-07-14; the arXiv paper followed on 09-22.

← Ledger