RSI-Anything: A Recursive Self-Improvement Framework

• RSI-Anything is a framework for recursive self-improvement that integrates predictive world models, feedback-driven policy revision, and conditional model adaptation to optimize coupled systems, enhancing efficiency in domains like code repair, GUI interaction, and laboratory planning.
• The framework distinguishes between external feedback and predictive modeling, allowing for selective allocation of resources to evaluate only the most promising candidates, which is beneficial in expensive feedback environments such as code execution and laboratory experiments.
• Applications of RSI-Anything, including code repair and laboratory planning, have shown significant improvements in success rates and operational efficiency, such as enhancing code repair success by over 5% and reducing errors in GUI interactions by 49% to essentially zero.

RSI-Anything is a domain-general framework for recursive self-improvement that combines predictive world models, feedback-driven policy or workflow revision, conditional model adaptation, and compatibility-based retention of updated policy–model pairs. Developed on JEPA-Anything, it treats recursive improvement as optimization of a coupled system $(\theta_k,\phi_{d,k})$, where $\theta_k$ denotes a policy or workflow and $\phi_{d,k}$ denotes a domain-specific predictive world model. The framework is evaluated in code repair, graphical-user-interface interaction, laboratory planning, reaction optimization, and an immune-science workflow (Cui et al., 9 Oct 2026).

1. Conceptual foundations and scope

Recursive self-improvement in RSI-Anything is defined as the repeated conversion of interaction experience into persistent changes in how an agent acts and predicts consequences. Conventional systems generate candidate actions, execute or evaluate them, use feedback to update a policy, workflow, memory, verifier, or code, and retain the resulting system for later cycles. RSI-Anything adds a predictive layer: a world model estimates the consequences of candidate actions before external execution, enabling selective allocation of costly feedback.

The framework addresses two coupled problems. First, external feedback may be expensive in code execution, GUI interaction, laboratory protocols, chemical experiments, and biological measurements. Second, a world model trained under an earlier policy may become stale when policy updates alter the distribution of proposed actions, visited states, and observed trajectories. Consequently, world-model adaptation cannot be treated as independent of policy evolution.

The retained object is therefore a policy–world-model pair:

$(\theta_k,\phi_{d,k}),$

where $\theta_k$ may be a neural policy, code generator, action-selection rule, or persistent agent workflow, and $\phi_{d,k}$ is the predictive model for domain $d$. Recursive cycles may end without changing either component; the cycle index counts attempted update-and-acceptance processes rather than only successful revisions.

RSI-Anything differs from frameworks that optimize only prompts, policies, or static workflows. The broader recursive-self-improvement literature includes executable harness search around frozen models, such as VideoHarness-RSI (Xu et al., 25 Aug 2026), and second-order diagnosis of workflow failures, such as SO-RSI (Jiang et al., 5 Oct 2026). RSI-Anything instead makes predictive modeling and policy–model compatibility central components of the recursive loop.

2. Task formulation and predictive world models

For domain $d$, RSI-Anything represents a task as

$$\mathcal{T}_d= (\mathcal{X}_d,\mathcal{S}_d,\mathcal{A}_d,\mathcal{W}_d, \mathcal{V}_d,\mathcal{M}_d,\mathcal{C}_d,\mathcal{B}_d).$$

The components comprise observations and context $\mathcal{X}_d$, task states $\theta_k$0, actions or interventions $\theta_k$1, the world-model adapter $\theta_k$2, a verifier or evaluator $\theta_k$3, evaluation metrics $\theta_k$4, operational, validity, and safety constraints $\theta_k$5, and computation, simulation, and external-interaction budgets $\theta_k$6.

The framework has two temporal scales. The action-step index $\theta_k$7 describes decisions within a task, while the recursive-cycle index $\theta_k$8 describes policy and world-model update attempts. At an action step, the current state and domain context are encoded as

$\theta_k$9

and a candidate action is used to predict a future latent representation:

$\phi_{d,k}$0

A domain-specific readout produces either a deterministic outcome,

$\phi_{d,k}$1

or a probabilistic prediction,

$\phi_{d,k}$2

The predicted outcome may represent program execution behavior, GUI-action success, laboratory-protocol consequences, reaction yield, or biological-state changes following an intervention. In code repair, the world model predicts five execution-result classes; in chemistry, it predicts a yield distribution.

RSI-Anything is built on JEPA-Anything, a cross-domain joint-embedding predictive architecture. Its implementation uses orthogonal predictive factorization (OPF), which decomposes target representations into complementary factors, predicts those factors separately, and recombines them into a complete latent prediction. A generic prediction objective is

$\phi_{d,k}$3

where $\phi_{d,k}$4 is an adaptation dataset and $\phi_{d,k}$5 is a domain-appropriate prediction loss.

For multistep tasks, the system can propagate imagined trajectories under policy $\phi_{d,k}$6:

$\phi_{d,k}$7

These trajectories remain hypothetical until validated externally. RSI-Anything therefore distinguishes prediction from observation rather than treating simulated outcomes as automatically equivalent to measured feedback.

3. Prediction-guided action and data selection

At state $\phi_{d,k}$8, the current policy proposes a candidate set

$\phi_{d,k}$9

The world model predicts each candidate’s consequences, after which the verifier removes invalid or unsafe candidates. Feasible candidates are ranked with a domain-specific utility:

$(\theta_k,\phi_{d,k}),$0

where $(\theta_k,\phi_{d,k}),$1 are predicted utility components, $(\theta_k,\phi_{d,k}),$2 are domain-specific weights, and $(\theta_k,\phi_{d,k}),$3 is an optional uncertainty penalty. When enabled,

$(\theta_k,\phi_{d,k}),$4

with $(\theta_k,\phi_{d,k}),$5 aggregating uncertainty estimates.

The selection mechanism can be utility-based or information-seeking. Utility-based selection chooses candidates with high predicted value. Information-seeking selection may deliberately choose uncertain candidates because their external outcomes can improve the world model. Both strategies are constrained by the remaining external-evaluation budget.

The framework distinguishes four evidence levels:

• E0: deterministic rule checks;
• E1: learned-model predictions and rollouts;
• E2: specialized simulators or higher-fidelity calculations;
• E3: external measurements or experiments.

Imagined trajectories are admitted as learning data only when they satisfy validity, utility, uncertainty, and provenance requirements. An admission indicator is defined as

$(\theta_k,\phi_{d,k}),$6

If uncertainty filtering is enabled,

$(\theta_k,\phi_{d,k}),$7

The resulting simulated dataset is

$(\theta_k,\phi_{d,k}),$8

This separation prevents a candidate from being excluded from external evaluation merely because it is uncertain. An uncertain candidate may be valuable precisely because executing it supplies information, even though its imagined trajectory is not trusted for model training.

External execution produces prediction–observation pairs stored in a feedback history $(\theta_k,\phi_{d,k}),$9. Separate subsets may be used for policy improvement, world-model adaptation, imagined-data training, component gates, and compatibility evaluation. The policy update is represented as

$\theta_k$0

The code-repair experiment sets $\theta_k$1, holding the coder fixed while evaluating predictive selection and world-model adaptation.

4. Conditional adaptation and retention control

World-model adaptation is conditional rather than automatic. The system requires evidence that the current model is unreliable, sufficient adaptation data, and adequate remaining budget. Reliability is assessed using predictions attributable to the current model version:

$\theta_k$2

An adaptation trigger is

$\theta_k$3

An “insufficient” reliability assessment does not automatically trigger adaptation. It indicates inadequate coverage rather than evidence that the model is reliable.

The proposed world-model update is

$\theta_k$4

The system evaluates policy and world-model updates independently before testing the assembled pair. Component checks are written as

$\theta_k$5

and

$\theta_k$6

Only components that pass their individual checks are assembled. The resulting pair is then evaluated using a compatibility gate:

$\theta_k$7

The final retention rule is

$\theta_k$8

The pair check is necessary because isolated component improvements may be incompatible. An updated policy may produce actions outside the predictor’s reliable range, while an updated predictor may rank actions in a way that harms the updated policy. A policy may also exploit weaknesses in the world model. The pair gate evaluates the actual decision system rather than relying solely on component-level metrics.

This retention structure is related to broader work on reliable recursive evaluation. REUSE provides statistical control of false promotions when an adaptive search repeatedly reuses a protected benchmark (Sun et al., 27 Sep 2026). RSI-Anything’s component and pair checks address compatibility and update safety, while REUSE addresses the statistical validity of deciding that an update is genuinely beneficial.

5. Applications and empirical findings

Code repair

The code-repair experiment holds the policy fixed using ReasonFlux-Coder-7B and compares world-model configurations. Four revisions are proposed per round, one is executed, and the benchmark contains 200 CodeContests tasks with up to three repair rounds.

Method	Hidden all-pass	Online calls per task
Random-$\theta_k$9	26.0%	3.72
Static verifier	26.3%	3.72
Fixed world model	29.2%	3.71
Adapted world model	31.2%	3.72

The fixed predictive model increases hidden success from 26.3% to 29.2% without materially increasing online test usage. World-model adaptation raises success further to 31.2%. Candidate-ranking accuracy improves from 66.5% to 72.8%, while the five-class Brier score decreases from 0.488 to 0.358. The experiment isolates predictive selection and adaptation rather than demonstrating simultaneous policy–model co-evolution.

GUI interaction

The GUI evaluation uses MiniWoB++ click-checkboxes and OSWorld. On click-checkboxes, success rises from 51% for the baseline to 79% after adaptation and 99% for the final RSI-Anything configuration. Mean actions per episode fall from 5.30 to 3.21, a 39.4% reduction. Repeated clicks decrease from 237 to 0, and episodes with incorrect submissions decrease from 9 to 1.

An intermediate configuration uses 2.93 actions per episode but produces 21 incorrect-submission episodes. This result illustrates the purpose of pair-level retention: minimizing actions is insufficient if efficiency is achieved at the expense of correctness. On OSWorld, RSI-Anything improves success by 18.4% relative to the initial agent and by 5.5% relative to RLAnything.

Laboratory planning

The laboratory-planning benchmark is LabHorizon F200, with 200 evaluation tasks, a 4B planning model, and a JEPA-Anything-based protocol model. Five successive policy versions receive the following scores:

$\phi_{d,k}$0

The total improvement is 24.21 percentage points. The final RSI-Anything score is 46.01%, compared with 45.32% for the Actor–Simulator–Selector reference. At Level 1, RSI-Anything achieves 0.700 accuracy versus 0.665 for the reference.

Reaction optimization

The chemistry adapter combines ReactionT5 with JEPA-Anything to predict a yield distribution

$\phi_{d,k}$1

where $\phi_{d,k}$2 is a reaction specification, $\phi_{d,k}$3 is a candidate condition vector, and $\phi_{d,k}$4 is reaction yield.

In a screen varying a Pd–NHC–DMS precatalyst, base, and solvent, RSI-Anything reaches 93.0% after 20 experiments, compared with 87.5% for Bayesian optimization. In a second screen involving Pd(OAc)$\phi_{d,k}$5, ligand, base, and solvent, mean best observed yield under a 40-experiment budget is 96.15% for RSI-Anything and 82.16% for Bayesian optimization.

These results concern two Suzuki–Miyaura screens and therefore do not establish universal superiority across reaction families or laboratory settings.

Immune-science workflow

The end-to-end immune-science application combines workflow evolution, predictive modeling, analysis, and therapeutic-hypothesis generation. The system represents cellular states, tissue immune context, and individual-level context, while modeling interventions such as cytokines, genetic perturbations, and candidate therapies.

An agent called Agent Genesis searches scientific-agent workflows and model configurations. A retained workflow constructs an action-conditioned immune world model, which is then used to simulate interventions, rank candidates, revise analyses, perform self-correction, and prioritize therapeutic hypotheses. The resulting prospective nomination combines SIRP$\phi_{d,k}$6 inhibition with IL-36$\phi_{d,k}$7-axis modulation. The hypothesis remains experimentally unvalidated.

6. Limitations, interpretation, and future directions

RSI-Anything does not eliminate the need for external feedback. It reallocates limited feedback by predicting candidate consequences and selecting which actions merit execution. The method remains vulnerable to biased or stale world models. A predictor can deprioritize valuable actions, creating a feedback loop in which it receives little evidence about the actions it systematically misses.

Selective feedback also limits the scope of reliability assessments. A world model may be reliable over the observed candidate range but inaccurate elsewhere. The framework distinguishes insufficient coverage from reliability, but the distinction does not itself solve the coverage problem.

The acceptance mechanism depends on trustworthy verifiers, held-out data, external tests, or domain-specific metrics. In physical laboratory and biological settings, benchmark performance does not guarantee successful or safe operation. Laboratory planning is evaluated on a benchmark rather than through complete execution on real equipment, and the immune-science therapeutic nomination awaits experimental testing.

The evidence for mutual policy–model co-evolution is also incomplete. The code-repair experiment holds the coder fixed, while the immune-science downstream analysis fixes world-model parameters. The paper identifies joint multiround evaluation as necessary to establish sustained mutual improvement of both components within one task.

Other limitations include adaptation cost, dependence on domain-specific prediction and reliability models, possible exploitation of predictor weaknesses, and the narrow scope of the reported chemistry experiments. Exact transfer to unseen sensors, environments, reaction families, or biological systems is not established.

The framework suggests several directions for further research: statistically controlled promotion of updates, richer uncertainty and coverage estimation, active exploration of poorly modeled action regions, safer pair gates for physical interventions, and recursive improvement of the proposer and evaluator themselves. The last direction introduces additional risks, because changing the evaluator can create reward-hacking opportunities. A general RSI system would therefore require protected evaluation, explicit error budgets, adversarial auditing, and separation between proposal generation and certification.

RSI-Anything’s principal methodological contribution is the treatment of recursive improvement as a coupled predictive-control problem:

$\phi_{d,k}$8

Its empirical results indicate that predictive selection can improve outcomes under fixed external-feedback budgets, and that conditional world-model adaptation can further improve selection when policies or workflows change. The framework does not establish unrestricted autonomous self-improvement; rather, it defines a structured mechanism for retaining policy–world-model updates when their isolated performance, compatibility, validity, and operational constraints are jointly satisfied.