First experimental evidence of recursive self-improvement (RSI)

Read on X Scrolly thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079778793042425" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079780575617364" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079782463005113" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079784421765620" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079786573484413" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079788767064448" dir="auto"> 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) thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079790625108243" dir="auto"> 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 (7/7) weco.ai/blog/first-evi… thread#showTweet" data-screenname="zhengyaojiang" data-tweet="2077079792118338027" dir="auto"> 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! • • •

Stay in touch and get notified when new unrolls are available from this author!

Twitter may remove this content at anytime! Save it as PDF for later use!

Practice here first or read more on our help page !

This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story .

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Make a small donation by buying us coffee ($5) or help with server cost ($10)

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Follow Us! Post Share Send Email! × Email the whole thread instead of just a link!