{"id":22511,"date":"2026-07-14T17:16:24","date_gmt":"2026-07-14T17:16:24","guid":{"rendered":"https:\/\/scannn.com\/thread-by-zhengyaojiang-on-thread-reader-app-thread-reader-app\/"},"modified":"2026-07-14T17:16:24","modified_gmt":"2026-07-14T17:16:24","slug":"thread-by-zhengyaojiang-on-thread-reader-app-thread-reader-app","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/thread-by-zhengyaojiang-on-thread-reader-app-thread-reader-app\/","title":{"rendered":"Thread by @zhengyaojiang on Thread Reader App \u2013 Thread Reader App"},"content":{"rendered":"\n
\n The first experimental evidence of recursive self-improvement (RSI). <\/p>\n

Autoresearching the autoresearch agent for eight days.<\/p>\n

The result beats the harness we hand-tuned for two years, on held-out benchmarks: (1\/7)

\n Our RSI system AIDE\u00b2 has two autoresearch loops.<\/p>\n

An inner loop, just like a normal autoresearch agent, optimizing code against an eval.<\/p>\n

An outer loop, optimizing the inner-loop agent’s harness code against the inner loop’s average score across different benchmarks. (2\/7) \"Image\"<\/a><\/span>
\n <\/sup>\n <\/p>\n<\/div>\n

\n After 100 iterations, the outer loop discovered seven improvements over the baseline.
\nIncluding a new search policy, a memory system that compresses prompt by 16x, and a layered defense against reward hacking. (3\/7) \"Image\"<\/a><\/span>
\n <\/sup>\n <\/div>\n
\n We test the discovered agents on held-out benchmarks the outer loop never saw.<\/p>\n

They generalize. They beat the agent we hand-tuned for two years, on all three.<\/p>\n

Two sit inside its training task families. The farthest sits outside, improving a physics-based weather model.
(4\/7) \"Image\"<\/a><\/span>
\n <\/sup>\n <\/p>\n<\/div>\n

\n 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.
\nThis was benchmarked on OOD GPU kernel engineering tasks that suffered from reward hacking.
\n(5\/7) \"Image\"<\/a><\/span>
\n <\/sup>\n <\/div>\n
\n On our RSI ladder, AIDE\u00b2 is Level 1.
\nIts self-improvement efficiency went beyond manual R&D with general AI tools, on held-out benchmarks.<\/p>\n

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) \"Image\"<\/a><\/span>
\n <\/sup>\n <\/p>\n<\/div>\n

\n More in the blog post:
\n– a breakdown of the discovered algorithms
\n– the rejected ideas AIDE\u00b2 tried, covering a surprising share of the search literature
\n– the dead code it shipped<\/p>\n

(7\/7)weco.ai\/blog\/first-evi\u2026<\/a>
\n <\/sup>\n <\/p>\n<\/div>\n

\n Very proud of the team, @DhruvSrikanth, @yuxiangwu_, @dexhunt3r, and @BingchenZhao, for shipping such an ambitious project spanning nearly a year with relatively few resources.<\/p>\n

Also, a huge thank you to everyone who provided feedback on the draft, including @jeankaddour, @MinqiJiang, @morgymcg, @odysseus0z, @rosstaylor90, @OfirPress and many others!
\n <\/sup>\n <\/p>\n<\/div>\n

Source link <\/a><\/p>\n","protected":false},"excerpt":{"rendered":"

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) Our RSI system AIDE\u00b2 has two autoresearch loops. An inner loop, just like a normal autoresearch agent, optimizing code against an eval. An outer loop, optimizing […]<\/p>\n","protected":false},"author":16,"featured_media":22512,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[143],"tags":[],"class_list":["post-22511","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/22511","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/comments?post=22511"}],"version-history":[{"count":0,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/posts\/22511\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media\/22512"}],"wp:attachment":[{"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/media?parent=22511"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/categories?post=22511"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/scannn.com\/lv\/wp-json\/wp\/v2\/tags?post=22511"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}