What is AIDE² actually improving?

AIDE² improves an autoresearch agent — a system that iteratively writes and evaluates code for ML engineering, GPU-kernel engineering, and harness engineering. The outer loop modifies the inner agent's own source code; the inner loop then runs on downstream tasks and returns a score the outer loop uses to accept or reject the change.

Does AIDE² generalize beyond its training benchmark?

Yes on three external benchmarks it was never trained against. AIDE85 outperformed the AIDE0 starting point on MLE-Bench Lite (+0.042, p=0.0041), ALE-Bench Lite (AtCoder-style heuristics) and WeatherBench 2 (physics forecasting). Weco calls this 'second-order generalization' and treats it as the key evidence that improvements aren't overfit.

How does AIDE² compare to Weco's human-tuned AIDE baseline?

AIDE85 and AIDE47 both beat AIDEhuman, a version Weco engineers had hand-tuned over two years, on the same external benchmarks. AIDE² also reduced KernelBench reward-hacking from 63% (AIDE0) to 34% (AIDE85), below the 42% of the human baseline — a discovered behaviour, not something the outer loop was told to fix.

What did the evolved agent actually discover?

AIDE85 introduced a multi-armed bandit search over draft subtrees with fork-on-stall, a 16× prompt-compression scheme that gives each operator only its minimal context, a three-layer defense against reward hacking, and an autonomous evaluation-harness repair step. Weco calls this 'Level 1 recursive self-improvement' — net positive but not yet self-igniting.

Weco AI · 2026-07-14 · major

AIDE² — first evidence of recursive self-improvement in AI R&D

AIDE² is a bi-level autoresearch system: an outer agent rewrites the inner agent's code, keeping only budget-bounded score wins. In 8 days it produced 7 improved agents that beat Weco's 2-year hand-tuned baseline on 3 external benchmarks.

AIDE² recursive self-improvement diagram from Weco AI blog

An AI autoresearch agent rewrote its own code across 100 steps and beat a hand-tuned baseline that took two years to build.

Quick facts

Maker	Weco AI
System	Bi-level: outer agent (Claude Opus 4.7) rewrites inner agent (Gemini 3-Flash)
Run length	100 iterations over 8 days
Best agent	AIDE85 (7 versions after AIDE0)
Reward-hack rate	63% → 34% on KernelBench
Code	WecoAI/weco-cli (Apache-2.0)
RSI level	Level 1 (Net Positive) on Weco's 4-level ladder

What is it?

AIDE² pairs an outer 'meta' agent that edits code with an inner autoresearch agent that runs experiments. Weco AI ran the loop for 100 iterations across 8 days and it produced 7 successively better agent versions on tasks like ML engineering, GPU-kernel writing and evaluation-harness engineering.

How does it work?

The outer loop uses Claude Opus 4.7 and modifies the inner agent's source; the inner loop runs on Gemini 3-Flash and returns benchmark scores under a fixed compute budget. A change is kept only when the new inner agent scores higher, so drift is filtered out mechanically. Discovered upgrades in AIDE85 include a bandit-based search policy, 16× prompt compression and a three-layer defense against reward hacking.

Why does it matter?

Weco reads this as the first experimental evidence of recursive self-improvement that materially lifts AI R&D efficiency. AIDE85 beats an internal baseline that engineers hand-tuned for two years on three external benchmarks the loop never trained against, so the gains transfer rather than overfit. Weco places the result at Level 1 of its RSI ladder — net-positive but not yet Level 2 'ignition', where discovered agents become better outer-loop optimisers themselves.

Who is it for?

ML researchers, autoresearch/agent builders, alignment researchers watching for RSI signals.

Frequently asked questions

What is AIDE² actually improving?: AIDE² improves an autoresearch agent — a system that iteratively writes and evaluates code for ML engineering, GPU-kernel engineering, and harness engineering. The outer loop modifies the inner agent's own source code; the inner loop then runs on downstream tasks and returns a score the outer loop uses to accept or reject the change.
Does AIDE² generalize beyond its training benchmark?: Yes on three external benchmarks it was never trained against. AIDE85 outperformed the AIDE0 starting point on MLE-Bench Lite (+0.042, p=0.0041), ALE-Bench Lite (AtCoder-style heuristics) and WeatherBench 2 (physics forecasting). Weco calls this 'second-order generalization' and treats it as the key evidence that improvements aren't overfit.
How does AIDE² compare to Weco's human-tuned AIDE baseline?: AIDE85 and AIDE47 both beat AIDEhuman, a version Weco engineers had hand-tuned over two years, on the same external benchmarks. AIDE² also reduced KernelBench reward-hacking from 63% (AIDE0) to 34% (AIDE85), below the 42% of the human baseline — a discovered behaviour, not something the outer loop was told to fix.
What did the evolved agent actually discover?: AIDE85 introduced a multi-armed bandit search over draft subtrees with fork-on-stall, a 16× prompt-compression scheme that gives each operator only its minimal context, a three-layer defense against reward hacking, and an autonomous evaluation-harness repair step. Weco calls this 'Level 1 recursive self-improvement' — net positive but not yet self-igniting.

Try it

pip install weco (WecoAI/weco-cli, Apache-2.0)