RSI-Anything Announced: Recursive Self-Improvement Guided by World Models

Ling Yang (@LingYang_PU) announced RSI-Anything in an X post dated October 12, 2026, around 11:16 AM KST. The proposal asks whether AI can improve itself by predicting the consequences of its actions: built on JEPA-Anything, a world model guides action selection and learning, while feedback improves the policy or workflow and, when needed, the world model itself.

Image source: Ling Yang (@LingYang_PU) X post attached video

According to the author, the core loop works as follows: the world model guides action choices and learning direction, the resulting feedback improves the policy or workflow (and updates the world model when needed), and only validated updates shape the next round of experience. No paper, repository, or blog link was included in the acquisition bundle, so every figure below is confirmed only as the author's claim in that single post, with no independent cross-verification.

The structure as announced and its reported scope

The announcement says results were reported across four areas: code, GUI interaction, laboratory planning, and reaction optimization. The post lists the figures together but does not include measurement conditions, datasets, or evaluation protocols.

Structurally, the author presents four elements: world-model guidance, policy-or-workflow improvement, world-model updates when needed, and carry-forward of validated updates only. The specific validation criteria are not stated in the post.

Reported per-domain figures

All figures below are relayed without independent verification, exactly as the narrow supported claims in the announcement allow.

• Code-repair success: the author reports a rise from 26.3% to 31.2% with the coder fixed.
• GUI agent tasks: the author reports 39.4% fewer actions alongside higher completion.
• Lab-planning scores: the author reports a rise from 21.80% to 46.01% over five policy versions.
• Mean best observed reaction yield: the author reports 96.15% versus 82.16%, but the comparator label is truncated at "for Bayesia…" in the source text, so the full comparator and conditions cannot be confirmed.

A "40-experiment budget" mentioned in replies is a third-party remark, not part of the announcement, and is not adopted as fact. This proposal should also not be confused with separate same-day work such as Sakana AI's MASS.

Confirmed limits and how to read this

The factual status here is "announcement confirmed as the author's claim," not "performance independently verified." There is no paper, repository, or blog link in the bundle, and auxiliary web search was skipped on budget exhaustion, so no cross-check exists. In particular, the reaction-yield comparison sentence is cut off after the comparator, so it must not be cited as a completed head-to-head fact.

This article therefore documents the existence of the RSI-Anything proposal and the structure and figures its author presented. Judgments about reproducibility or general performance must wait for follow-up materials (paper, code, evaluation conditions).

Sources

• Ling Yang (@LingYang_PU) X post: Introducing RSI-Anything: Unlocking Recursive Self-Improvement with World Models