Skip to content
ThreadOut
Zhengyao Jiang@zhengyaojiang · Jul 14, 2026

The first experimental evidence of recursive self-improvement (RSI). Autoresearching the autoresearch agent for eight…

8 tweets1 min read♥ 9.6Koriginal

Summary

Researchers demonstrated the first experimental evidence of recursive self-improvement (RSI) using AIDE², an autoresearch system with nested optimization loops that discovered improvements to its own code over 8 days, outperforming manually hand-tuned baselines on held-out benchmarks including physics-based tasks.

Summarized by ThreadOut AI from the full thread. May miss nuance — read the thread below.

  1. #1

    The first experimental evidence of recursive self-improvement (RSI). Autoresearching the autoresearch agent for eight days. The result beats the harness we hand-tuned for two years, on held-out benchmarks: 🧵(1/7)

  2. #2

    Our RSI system AIDE² has two autoresearch loops. An inner loop, just like a normal autoresearch agent, optimizing code against an eval. An outer loop, optimizing the inner-loop agent's harness code against the inner loop's average score across different benchmarks. (2/7)

  3. #3

    After 100 iterations, the outer loop discovered seven improvements over the baseline. Including a new search policy, a memory system that compresses prompt by 16x, and a layered defense against reward hacking. (3/7)

  4. #4

    We test the discovered agents on held-out benchmarks the outer loop never saw. They generalize. They beat the agent we hand-tuned for two years, on all three. Two sit inside its training task families. The farthest sits outside, improving a physics-based weather model. (4/7)

  5. #5

    We also see an emergent phenomenon where the outer loop pushes the inner-loop agent's reward hacking rate lower, with a combination of prompting and rule-based checks. This was benchmarked on OOD GPU kernel engineering tasks that suffered from reward hacking. (5/7)

  6. #6

    On our RSI ladder, AIDE² is Level 1. Its self-improvement efficiency went beyond manual R&D with general AI tools, on held-out benchmarks. We also tested Level 2, whether the improved inner agent makes a better outer loop. Results are mixed, and we do not claim ignition. (6/7)

  7. #7

    More in the blog post: - a breakdown of the discovered algorithms - the rejected ideas AIDE² tried, covering a surprising share of the search literature - the dead code it shipped weco.ai/blog/first-evi… (7/7)

  8. #8

    Very proud of the team, @DhruvSrikanth, @yuxiangwu_, @dexhunt3r, and @BingchenZhao, for shipping such an ambitious project spanning nearly a year with relatively few resources. Also, a huge thank you to everyone who provided feedback on the draft, including @jeankaddour, @MinqiJiang, @morgymcg, @odysseus0z, @rosstaylor90, @OfirPress and many others!