RSI-Anything: Unlocking Recursive Self-Improvement in AI with World Models

Why Recursive Self-Improvement Has Long Remained Theoretical

Recursive Self-Improvement (RSI) has always been one of the most compelling goals in AI research: enabling a system to convert every interaction with its environment into sustained capability gains. In practice, however, this path faces two structural barriers.

The first is the high cost of external feedback. Whether in code testing, scientific experimentation, or real-world user interactions, obtaining reliable feedback signals typically demands significant computational, time, or physical resources. The second barrier is more subtle — when the policy itself is continuously evolving, the world model the system relies on to predict outcomes may gradually lose accuracy. In other words, the more the model improves, the more its judgments about the consequences of its own actions can drift from reality.

RSI-Anything is a general-purpose framework designed specifically to address these two pain points. Built on top of JEPA-Anything, its core idea is to use a predictive world model to guide two things: which actions to take, and which data to select for learning.

JEPA (Joint Embedding Predictive Architecture) is a class of world model architectures proposed by Yann LeCun's team. Unlike traditional generative models, JEPA does not attempt to predict the future in pixel or raw input space — instead, it predicts target representations in an abstract embedding space. The advantage of this design is that the prediction task focuses on semantic-level structural information, avoiding futile modeling of unpredictable details (such as texture and noise), which reduces computational cost and improves representation quality. JEPA-Anything is a generalized extension of this architecture, aimed at adapting it to multiple modalities and task types. RSI-Anything builds on this foundation by embedding a JEPA-style predictive world model into the self-improvement loop, enabling agents to "simulate" outcomes in embedding space before acting. This allows them to filter for high-value actions and training data without paying the cost of real environment interactions for every decision.

How the Framework Makes Improvement "Persistently Retainable"

The operational logic of RSI-Anything can be broken down into several interlocking components. Observed feedback simultaneously drives two improvement pathways: optimization of the policy or workflow itself, and conditional adaptation of the world model. That is, the system is not only learning how to act — it is also learning to more accurately predict the consequences of its actions.

The key innovation lies in the "retention mechanism." The framework introduces component checks and compatibility checks to determine whether an updated policy–model pair is worth retaining. This step serves as a quality gate for recursive improvement — only updates that pass validation are solidified into persistent system capabilities, preventing the model from diverging or degrading through self-iteration.

This design directly addresses the two barriers mentioned earlier: predictive selection reduces dependence on expensive external feedback, while conditional world model adaptation combined with retention checks mitigates the model accuracy drift caused by policy evolution.

What the Cross-Domain Experimental Results Reveal

The most compelling aspect of RSI-Anything is that it demonstrates substantial gains across multiple highly diverse domains, rather than being confined to a single benchmark.

Code Repair

With the coder held fixed, code repair success rates show a clear stepwise improvement: 26.3% with a static validator, rising to 29.2% after introducing predictive selection, and reaching 31.2% after world model adaptation. Importantly, these gains were achieved with comparable test resource consumption, indicating that the improvements stem from smarter selection rather than simply stacking more compute.

GUI Operations

In GUI task evaluations, RSI-Anything not only improved task completion rates but also reduced the average number of operational steps by 39.4%. For automation agents, fewer actions mean higher efficiency and lower cumulative risk of error.

Scientific Planning and Reaction Optimization

Over five consecutive policy iterations, laboratory-planning scores more than doubled — from 21.80% to 46.01% — vividly demonstrating the cumulative effect of recursive improvement.

In reaction optimization tasks, RSI-Anything achieved an average best-observed yield of 96.15% within the same budget of 40 experiments, significantly outperforming Bayesian Optimization at 82.16%. This is a noteworthy comparison, as Bayesian Optimization has long been a strong baseline in experimental design.

Bayesian Optimization is a sequential experimental design method for expensive black-box functions, widely used in chemical synthesis, materials science, and hyperparameter tuning. Its core idea is to use a probabilistic surrogate model (typically a Gaussian process) to build a confidence estimate of the objective function's shape, then use an acquisition function (such as Expected Improvement, EI) to balance "exploring unknown regions" versus "exploiting known high-performing regions," thereby approaching the optimum with as few experiments as possible. Since each real experiment (e.g., a chemical reaction) is costly, Bayesian Optimization has long been regarded as the gold-standard baseline for such tasks. RSI-Anything surpassing it with a yield of 96.15% versus 82.16% under the same 40-experiment budget implies that the structured priors learned by the predictive world model are, in some sense, better at capturing the underlying patterns of reaction optimization tasks than the statistical priors of Gaussian processes — a conclusion that, if validated across broader chemical spaces, would carry significant practical value.

From Workflow Evolution to Scientific Hypothesis Generation

The research also presents an end-to-end immunoscience application case that chains together workflow evolution, predictive modeling, and therapeutic hypothesis generation. The significance of this type of application is that it demonstrates recursive self-improvement is not merely about boosting an isolated metric — it has the potential to drive automation across an entire research pipeline, from experimental planning and outcome prediction to the generation of new hypotheses.

This also distinguishes RSI-Anything from work that only scores on abstract benchmarks. What it aims to demonstrate is a transferable capability: a predictive world model, combined with validated and retained system updates, can jointly support continuous self-improvement across vastly different real-world environments.

Value and Points Worth Watching

The core contribution of RSI-Anything is translating "recursive self-improvement" — a largely theoretical goal — into an engineerable, cross-domain reusable set of mechanisms: predictive selection handles cost reduction, conditional model adaptation maintains accuracy, and component and compatibility checks ensure stable retention.

However, based on publicly available information, several questions remain worth monitoring. For instance: can the retention mechanism still prevent degradation over longer iteration chains? What is the computational overhead of world model adaptation? And can these gains be reproduced in more open-ended real-world scenarios? For practitioners focused on autonomous AI learning and long-term agent evolution, this work provides a framework paradigm well worth deeper investigation.

Background

Conditional adaptation differs fundamentally from unconditional continual learning. The core challenge in unconditional continual learning is "catastrophic forgetting": when a neural network updates its parameters on new data, it tends to overwrite knowledge acquired from previous data, leading to overall performance degradation. RSI-Anything's retention mechanism — by executing compatibility checks before updates are solidified — essentially plants an anti-degradation safeguard within the recursive improvement loop. Only when a new policy–model pair demonstrates in compatibility testing that it does not harm existing capabilities is the update written into the system. This design philosophy aligns in spirit with methods like Elastic Weight Consolidation (EWC) or Progressive Neural Networks (PNN) in the continual learning literature, but its implementation approach is closer to system-level component validation rather than parameter-level regularization constraints.