Recursive self-improvement (RSI) is the holy-grail narrative of AI agents, but it has always stumbled on two practical problems: real-world feedback is expensive, and world-model predictions drift out of reliability as the policy evolves. RSI-Anything, submitted to arXiv on Oct 9 (arXiv:2610.13912), offers an engineering-flavored answer: let the world model evolve together with the policy, and gate every upgrade with compatibility checks.

One loop, five domains

The core is a six-step cycle: Propose → Predict → Select → Evaluate → Learn → Retain. A policy or workflow proposes candidate actions; a JEPA world model predicts their consequences; the system picks which actions deserve real execution and external feedback. That feedback drives both policy improvement and conditional world-model adaptation, while component and compatibility checks decide whether the updated policy-model pair is retained for the next round.

The foundation is JEPA-Anything (arXiv:2609.20800), the same group's cross-domain predictive framework released in September — this paper is effectively a second floor on top of it. The same loop runs code repair, GUI interaction, laboratory planning, reaction optimization and immune science, swapping only a domain adapter per field.

What the numbers say

Code repair shows the mechanism best: with the coder fixed, a static verifier picks patches at 26.3% success; predictive selection lifts it to 29.2%, and world-model adaptation to 31.2% at comparable test usage. Extended to APPS Introductory (1,000 problems) and LiveCodeBench v6 (1,055 problems), final full-test success reaches 46.37% and 33.05% with a ReasonFlux-Coder-7B backbone.

On GUI, MiniWoB++ checkbox tasks go 51 → 79 → 99 completions across three retained versions; average actions drop from 5.30 to 3.21 and repeated clicks fall from 237 to zero. OSWorld desktop tasks improve 18.4% relative to the initial agent. Laboratory planning doubles from 21.80% to 46.01% across five policy versions. In Suzuki reaction optimization, the framework reaches 96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget, with every run finding the pool optimum by observation 10.

What matters — and what to stay calm about

Two things matter. First, the bottleneck of RSI was never "can it change" but "what if it changes for the worse": retention here is an explicit mechanism, gated by component and compatibility checks — effectively a rollback key for self-improvement. Second, Gen-Verse has shipped JEPA-Anything and RSI-Anything within three months; predictive world models are turning from paper concepts into reusable infrastructure, with a 402-star Apache-2.0 repo, HuggingFace datasets and an rsi CLI.

Caveats, plainly: all numbers are self-reported by the paper and official repo with no independent replication yet; 96.15% is a within-pool screening result, not "AI can do chemistry"; and the immune-science application is a workflow demo with prospective hypothesis nomination (SIRPα inhibition plus IL-36γ-axis modulation), not a validated therapy.

Paper: https://arxiv.org/abs/2610.13912
Code: https://github.com/Gen-Verse/RSI-Anything