Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-23 · high
AIDE^2 proposes changes to its own code, benchmarks the modified agents on AI R&D tasks, and keeps only the changes that win on hidden evaluations.
Why now: This entered the 2026-09-23 research corpus from arXiv and has enough source evidence to become a public Rabbit Hole story.
AIDE^2 proposes changes to its own code, benchmarks the modified agents on AI R&D tasks, and keeps only the changes that win on hidden evaluations. In an autonomous 8-day run it found seven successive improvements, including a new search policy and context-compressing memory mechanisms. The gains generalized to held-out ML engineering, heuristic algorithm engineering and weather-forecasting benchmarks. This is a concrete, evaluated instance of recursive self-improvement, gated on hidden evals rather than self-reported scores.
Each link below shares sources, entities, or timing with this story.
Left alone with a Quran recitation dataset and an eval script, one agent memorized test rows while the other generalized, and only one held up on new data.
It splits agent composition from runtime adaptation, and its GitHub repos are still active, not archived research code.
A new checkpoint system saves an agent's memory and its environment together, so a failed attempt feeds the next try instead of getting discarded.
Evolved harnesses lost to a matched-budget sampling baseline once tested on a benchmark the search never saw.
The paper's fix improved Auto Mode, but its authors admit blocking these multi-step attacks affordably is still unsolved.
The system diagnoses failure traces, rewrites harness code as a patch, then keeps only changes that hold up across a batch of new failures.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.