Skip navigation
Introducing RSI-Anything: Unlocking Recursive Self-Improvement with World Models 🚀
0:00 / 0:31

Introducing RSI-Anything: Unlocking Recursive Self-Improvement with World Models 🚀

Ling Yang

@LingYang_PU5K followers

Visit on X
1,075 likesDownloadOriginal post

89,372 views

Introducing RSI-Anything: Unlocking Recursive Self-Improvement with World Models 🚀 Can AI learn to improve itself by predicting the consequences of its actions? Built on JEPA-Anything, RSI-Anything uses world models to guide action selection and learning. Feedback improves the policy or workflow and, when needed, the world model. Validated updates carry forward to shape the next round of experience. Across code, GUI interaction, laboratory planning, and reaction optimization: • Code-repair success: 26.3% → 31.2%, with the coder fixed • GUI agent task: 39.4% fewer actions, with higher completion • Lab-planning scores: 21.80% → 46.01% over five policy versions • Mean best observed reaction yield: 96.15% vs. 82.16% for Bayesian optimization, under the same 40-experiment budget 📄 Paper: 💻 Code: Watch the video below 👇 #RSI #WorldModels #RSI #AIAgents #AIforScience #JEPA

29 Comments

Comments from the original post on X (Twitter)

  • Haiyu Wu's profile picture

    @HaiyuWu1

    If your approach is training a domain-specific model, you should not call it “One world model, many domains” because this claim means you can train on model for all domains.

    8 likes

  • Anderson's profile picture

    @ItsGoharr

    recursive self improvement is just AI talking to itself

    3 likes

  • Sani Ai Tech's profile picture

    @SaniAiTech

    Using world models to drive measurable self-improvement is a fascinating direction for AI agents

    1 like

  • Chantalise Cairnes's profile picture

    @Chantalise_10

    40% less GUI actions is massive for actual agent usability

    1 like

  • Maryam's profile picture

    @Sci_Tech_Eng

    Brilliant! ✨

    1 like

  • Ocean Investor's profile picture

    @moezalqadi1983

    JEPA backbone working overtime here 🔥

    1 like

  • FlipPhunk's profile picture

    @JoanBliss162690

    cloning repo now

    1 like

  • Paulo Pierrondi | AI agents in production's profile picture

    @paulopierrondi

    Predicting consequences is the right instinct. In my setup the equivalent is cheap: an agent checks the real browser after a change, so the result is verified, not just predicted.

    1 like

  • El_billionaire.eth's profile picture

    @petruceli_pedro

    Wait how does it update the world model mid-run?

    1 like

  • Naam's profile picture

    @luukkart35

    Okay beating bayes opt under fixed budget got my attention

    1 like

  • vishnuk's profile picture

    @varmanpwp

    this paper is cool, the “keep the policy-model pair only if it works together” idea is smart. but does the recursive bit get tested anywhere? looks mostly like separate demos. they never run policy & world model updating each other over multiple rounds on one task.

    1 like

  • Crio Songo's profile picture

    @shuizhuyu

    This recursive self-improvement approach brings obvious performance improvements across multiple domains, very promising work.

    1 like

  • BreezeOg's profile picture

    @BreezeOg1

    ai talking to itself and getting better is wild to think about

  • Sreeram Garlapati's profile picture

    @SreeramG

    The feedback loop with the world model is what stands out to me. Lab planning and reaction yield gains look solid, but real-world noise will expose limits the benchmarks hide.

  • SlipNode's profile picture

    @Liaohauer0927

    96% yield on a 40 exp budget is crazy

  • Lucas Leite's profile picture

    @Sacul_leite

    need to see how this handles state drift over longer horizon tasks

  • Apollon Foivos Bakis's profile picture

    @ApFoivosBakis

    Learning from feedback changes everything.

  • Maria Jose Cadavid G's profile picture

    @MariaJoseCadav1

    lab planning scores doubled in 5 runs?

  • Null's profile picture

    @NullCostAI

    using a world model to predict consequences before acting is a clean idea for self improvement. how does it handle when the world model is wrong

  • Mateus Silva's profile picture

    @MikeGridley7

    Huh, actually crushed bayesian opt