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

Abstract: Recursive self-improvement (RSI) aims to transform interaction experience into persistent improvements, yet external feedback is costly and world-model predictions may become unreliable as policies evolve. We introduce RSI-Anything, a domain-general framework built on JEPA-Anything that uses predictive world models to guide action and learning-data selection. Observed feedback drives policy or workflow improvement and conditional world-model adaptation, while component and compatibility checks govern the retention of updated policy-model pairs. Across diverse computational and scientific domains, RSI-Anything demonstrates substantial gains. In code repair, with the coder fixed, success increases from 26.3% with a static verifier to 29.2% through predictive selection and 31.2% after world-model adaptation, at comparable test usage. Graphical user interface (GUI) evaluations showed higher task completion, while average actions decrease by 39.4%. Across five successive policy versions, laboratory-planning scores more than double from 21.80% to 46.01%. In reaction optimization, RSI-Anything achieves 96.15% mean best observed yield versus 82.16% for Bayesian optimization under the same 40-experiment budget. An end-to-end immune-science application further connects workflow evolution and predictive modeling with therapeutic hypothesis generation. These results demonstrate how predictive world models and retained system updates enable recursive improvement across diverse environments.

• The paper presents a framework for recursive self-improvement (RSI) that combines policy updates and world-model predictions, enhancing accuracy and efficiency in learning across various domains.
• RSI-Anything demonstrates improved performance in code repair, GUI interactions, laboratory planning, and reaction optimization, with notable gains in hidden-test success and experimental efficiency.
• The framework includes adaptive mechanisms and retention checks, ensuring compatibility between policy updates and world models, crucial for sustained, reliable improvements.

Problem formulation and central thesis

“RSI-Anything: Unlocking Recursive Self-Improvement with World Models” (2610.13912) proposes a framework for recursive self-improvement (RSI) in which a policy and a predictive world model are treated as a coupled, persistent system. The central claim is that feedback-driven policy updates alone are insufficient for reliable RSI under constrained interaction budgets. A world model can reduce the cost of external evaluation by predicting candidate consequences, but its predictions may become unreliable when policy updates shift the candidate distribution. RSI-Anything therefore couples three operations: prediction-guided action and feedback selection, conditional adaptation of the world model, and retention of policy–model pairs through component-level and compatibility checks.

The framework builds on “JEPA-Anything: Learning Predictive Models across Different Worlds” (Cui et al., 17 Sep 2026), which provides a cross-domain predictive architecture based on joint-embedding prediction and orthogonal predictive factorization. RSI-Anything does not treat this predictive substrate as a static planning module. Instead, it embeds the predictor in a recursive update loop: predictions determine which candidates receive costly external evaluation; observed outcomes supervise policy or workflow revision and may trigger predictor adaptation; and acceptance tests determine whether the resulting policy–model pair governs the next cycle.

This formulation addresses a specific failure mode in recursive learning. A policy update changes the distribution of actions and states encountered by the world model, while a world-model update changes candidate rankings, imagined trajectories, and learning signals. Improving either component independently therefore does not establish improvement of the assembled system. RSI-Anything makes compatibility between the two components an explicit retention criterion.

Figure 1: RSI-Anything integrates cross-domain world-model prediction, feedback-driven policy or workflow revision, conditional predictor adaptation, and retained policy–model pairs.

Architecture and recursive update mechanism

For a domain $d$, RSI-Anything defines a task interface containing observations, states, actions, a world-model adapter, a verifier, metrics, constraints, and resource budgets. The accepted system at recursive cycle $k$ is represented by a policy configuration $\theta_k$ and a domain-specific world model $\phi_{d,k}$. The policy may consist of trainable parameters or a persistent workflow; the world model predicts action-conditioned outcomes or, where supported, multistep state transitions.

At each cycle, the current policy proposes a candidate set. The world model predicts candidate outcomes, and a verifier removes invalid or infeasible candidates before selection. A domain-specific utility function ranks the remaining candidates, optionally incorporating uncertainty penalties. Only selected candidates are externally executed or evaluated, producing prediction–observation pairs. This distinction is important: the model may predict every candidate, but external feedback is obtained only for the candidates chosen under the current selection rule.

The paper separates several evidence classes: deterministic checks, learned predictions and rollouts, specialized simulators or high-fidelity calculations, and external measurements. Imagined trajectories can contribute to policy learning only when they satisfy domain-specific validity, utility, uncertainty, and evidence-source criteria. Thus, RSI-Anything does not equate a model-generated trajectory with an observation; provenance is retained and training admission is separately controlled.

Policy updates and world-model adaptations are asynchronous. A policy update can be based on demonstrations, preferences, evaluated trajectories, or workflow search, depending on the domain. World-model adaptation is triggered only when reliability is assessed as inadequate, sufficient adaptation data are available, and the relevant budget permits the update. An insufficiently covered prediction range is not treated as evidence of reliability and does not automatically justify adaptation.

The retention mechanism has three levels. First, the candidate policy is evaluated against the current policy. Second, the candidate world model is evaluated against the current predictor. Third, the assembled candidate pair is assessed under matched conditions. If the pair fails compatibility checks, the entire current pair is retained even if one component has individually improved. This design makes persistence conditional on system-level behavior rather than on isolated component metrics.

Figure 2: Prediction-guided candidate selection supplies feedback for policy revision and conditional world-model adaptation, while component and compatibility checks determine which pair is retained.

The proposed mechanism is consequently more restrictive than simply accumulating interaction histories. A longer history does not constitute recursive improvement unless it changes the accepted policy or predictor. Conversely, a cycle may be counted even when neither component changes, because the framework distinguishes update attempts from successful revisions.

Code repair: predictive selection and model adaptation

The cleanest controlled test isolates the contribution of world-model-guided selection from policy capacity. The code-repair experiments hold the ReasonFlux-Coder-7B coder fixed and generate four repair candidates per round. The comparison uses 200 CodeContests tasks, three random seeds, and up to three feedback rounds. A static verifier, random selection, a fixed world model, and an adaptive world model are compared under similar online test usage.

The fixed-world-model configuration increases hidden-test all-pass success from $26.3\%$ for the capacity-matched static verifier to $29.2\%$, a paired gain of $2.83$ percentage points with a reported 95% interval of $[0.17, 5.50]$. Random selection obtains $26.0\%$. The predictive improvement is accompanied by better candidate evaluation: ranking accuracy rises from $66.5\%$ to $k$0, while the five-class Brier score decreases from $k$1 to $k$23.71$k$33.72$$for the static verifier. The implication is that the fixed predictor improves allocation of feedback rather than simply increasing the amount of feedback.</p> <p>World-model adaptation produces a further increase to$$k$4 hidden-test success, or two percentage points above fixed-world-model selection, with $k$5 online test calls per task. The coder remains fixed, so this comparison attributes the additional gain to predictor adaptation and its effect on subsequent candidate selection. Across APPS Introductory and LiveCodeBench v6, the predictive selection configuration achieves full-test success rates of $k$6 and $k$$7, respectively, under the same coder and candidate-generation protocol.</p> <p><img src="https://images.emergentmind.com/paper-images/2610-13912/figure2_code_repair_ppt_en.png" alt="Figure 3" title="" class="markdown-image" loading="lazy"></p> <p><p class="figure-caption">Figure 3: In code repair, predictive candidate ranking and feedback-driven world-model adaptation improve hidden-test success without increasing online test usage.</p></p> <p>The code-repair results provide the strongest evidence for the paper’s staged thesis: predictive selection yields an initial gain, and adapting the predictor from execution feedback yields a further gain. They do not, however, demonstrate unrestricted joint policy–model co-evolution, because the coder is intentionally held fixed in the principal ablation. This makes the causal interpretation clearer but leaves policy adaptation to other experiments.</p> <h2 class='paper-heading' id='gui-interaction-task-completion-and-behavioral-efficiency'>GUI interaction: task completion and behavioral efficiency</h2> <p>The GUI experiments evaluate MiniWoB++ click-checkboxes and <a href="https://www.emergentmind.com/topics/osworld" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">OSWorld</a>. Here, both the action policy and the world model are updated through interaction feedback, and the retained pair determines later actions. The MiniWoB++ diagnostics are particularly informative because they measure not only task success but also action count, repeated clicks, and incorrect submissions.</p> <p>For click-checkboxes, success increases from$$k$8 at baseline to $k$9 in the adapted configuration and $\theta_k$0 in the final RSI-Anything configuration. Mean actions per episode decline from $\theta_k$1 to $\theta_k$2, a $\theta_k$3 reduction. Repeated clicks decrease from 237 to zero, while episodes with incorrect submissions fall from nine to one. The adapted configuration uses fewer actions still—$\theta_k$4 per episode—but produces 21 incorrect-submission episodes. This contrast indicates that action minimization alone is not an adequate objective: the final retained configuration sacrifices a small amount of action-count efficiency relative to the adapted intermediate state while substantially improving submission reliability.

The World-based control reaches $\theta_k$5, $\theta_k$6, and $\theta_k$7 success across its corresponding configurations. On OSWorld, RSI-Anything improves success by $\theta_k$8 relative to the initial agent and by $\theta_k$9 relative to RLAnything. RLAnything itself improves by $12.3\% relative to the initial agent. These comparisons indicate that the reported gain is not reducible to generic online adaptation, although the paper does not provide a full factorial ablation separating policy-only and world-model-only updates on OSWorld.

Figure 4: GUI adaptation improves completion and reduces redundant actions, repeated clicks, and incorrect submissions across successive configurations.

The GUI results extend the code-repair findings from candidate ranking to sequential interaction. They also illustrate why pair retention matters: a policy that acts efficiently but submits incorrectly is inferior to a slightly less efficient policy that reliably completes the task.

Laboratory planning and successive policy versions

The LabHorizon evaluation tests whether improvements persist across successive policy versions in a physical-domain planning setting. The benchmark contains 200 Level-2 F200 tasks, where the Final Score combines action-sequence similarity with parameter accuracy. RSI-Anything follows five consecutive versions of a 4B policy generated within one self-improvement process.

The Final Score increases monotonically across the reported versions:

Policy version	Final Score
Initial	$\phi_{d,k}$0
Version 2	$\phi_{d,k}$1
Version 3	$\phi_{d,k}$2
Version 4	$\phi_{d,k}$3
Final	$\phi_{d,k}$4

The final policy improves over the initial version by $\phi_{d,k}$5 percentage points and exceeds the Actor–Simulator–Selector reference score of $\phi_{d,k}$6 by $\phi_{d,k}$7 percentage points. On Level-1 next-action selection, RSI-Anything obtains $\phi_{d,k}$8 accuracy versus $\phi_{d,k}$9 for the reference.

Figure 5: Five successive retained policy versions improve LabHorizon Level-2 planning from $26.3\%$0 to $26.3\%$1.

The monotonic sequence is consistent with persistent policy improvement rather than one-shot selection. Nevertheless, the benchmark evaluates protocol-aligned planning rather than execution on laboratory equipment. The result therefore establishes improvement in the evaluated planning representation, not reliable physical execution under real instrumentation constraints.

Reaction optimization under matched experimental budgets

The chemistry experiments assess whether prediction-guided selection improves experimental search when external measurements are expensive. RSI-Anything combines a ReactionT5-based reaction predictor with the JEPA-Anything predictive design and uses measured yields to select subsequent reaction conditions. Bayesian optimization serves as the comparator within two Suzuki–Miyaura screening protocols, with matched experiment counts.

In the first GC-analyzed screen, RSI-Anything reaches a best observed yield of $26.3\%$2 after 20 experiments, compared with $26.3\%$3 for Bayesian optimization, a gain of $26.3\%$4 percentage points. In the second flow LC–MS screen, RSI-Anything reaches a mean best observed yield of $26.3\%$5 after 40 experiments, versus $26.3\%$6 for Bayesian optimization, a gain of $26.3\%$7 percentage points.

Figure 6: At matched experiment counts, prediction- and feedback-guided reaction search obtains higher best observed yields than Bayesian optimization in two separate screens.

These results support the framework’s resource-allocation rationale: predicted yields are used to prioritize experiments, while measured yields update the evidence available for the next selection. The comparison is nonetheless narrow. It covers two Suzuki–Miyaura screens, and the reported metric is best observed yield rather than a broader measure of calibration, robustness, reproducibility, or transfer across reaction families. The gains therefore demonstrate performance in the reported screening regimes rather than general dominance over Bayesian optimization.

Immune-science workflow and hypothesis generation

The immune-science application applies RSI-Anything to a multistage scientific workflow. The construction stage evolves scientific-agent workflows and model configurations, selecting a workflow that produces an action-conditioned immune world model. The resulting model represents cellular, tissue, and individual immune states and predicts the consequences of explicit interventions. The downstream analysis stage uses the constructed predictor with fixed model parameters while the analysis workflow performs simulation, evidence integration, self-correction, intervention ranking, and hypothesis generation.

Figure 7: The immune-science application connects workflow evolution, multiscale immune-world-model construction, self-correcting analysis, and therapeutic hypothesis generation.

A notable feature is the separation between workflow evolution and downstream inference. The retained workflow governs model construction, while later analysis evolves candidate interventions and hypotheses without updating the world-model parameters. Self-correction removes screened cytokine-pair candidates, after which the integrated evidence and model analysis nominate a combination of SIRP$26.3\%$8 inhibition and IL-36$26.3\%$9-axis modulation.

This application demonstrates architectural composition rather than a controlled therapeutic efficacy result. The nominated intervention remains a prospective hypothesis awaiting experimental validation. Its significance for RSI-Anything is that the framework can connect persistent workflow selection, predictive modeling, and evidence-guided scientific reasoning within a single pipeline; it does not establish that the proposed therapy is biologically effective.

Relation to prior approaches

RSI-Anything extends several established lines of work. Verifier-based learning, preference optimization, reinforced self-training, and verbal reinforcement convert evaluated trajectories into policy-improvement signals (Cobbe et al., 2021, Gulcehre et al., 2023, Shinn et al., 2023). Coding systems such as CURE and self-improving coding agents retain validated revisions to code or agent procedures (Wang et al., 3 Jun 2025, Robeyns et al., 21 Apr 2025). Other systems evolve memories, curricula, environments, or complete agent implementations (Wang et al., 2023, Guo et al., 22 Dec 2025, Xia et al., 20 Nov 2025, Zhang et al., 29 May 2025).

The distinct contribution of RSI-Anything is to make predictive world models first-class participants in the recursive loop. World models have long supported planning and control (Ha et al., 2018, 2007.16770), while JEPA-style architectures provide latent predictive representations (Assran et al., 2023, Bardes et al., 2024, Assran et al., 11 Jun 2025). RSI-Anything specifies how these predictors should guide feedback acquisition, how their reliability should be assessed under distribution shift, and how compatibility with an evolving policy should determine retention.

The framework also differs from approaches that jointly optimize policy, reward, and environment, such as RLAnything (Wang et al., 2 Feb 2026). RSI-Anything does not simply allow all components to change simultaneously. It introduces asynchronous update schedules, provenance-sensitive data handling, reliability-triggered model adaptation, and a pair-level acceptance test. The result is a governance protocol for recursive updates rather than only a learning objective.

Limitations and open questions

The empirical evidence spans several domains, but the update scope differs substantially across them. Code repair tests world-model selection and adaptation with a fixed coder; laboratory planning evaluates successive policy versions but does not establish closed-loop physical execution; chemistry evaluates two reaction screens under fixed budgets; and the immune-science application freezes the world model during downstream analysis. Consequently, the paper demonstrates complementary components of RSI-Anything across domains, but does not yet establish sustained mutual improvement of both policy and world model in a single long-horizon task.

The reliability protocol is necessarily dependent on feedback coverage. Because candidates are selected before external evaluation, reliability estimates are limited to the observed candidate range and can be biased toward regions already favored by the current predictor. The paper explicitly treats insufficient coverage as distinct from reliability, but the experiments do not fully quantify how selection bias affects adaptation triggers or pair acceptance.

Several comparisons also leave methodological questions open. The GUI results do not fully isolate policy updates from world-model updates, and the chemistry experiments compare against Bayesian optimization only in two Suzuki–Miyaura settings. The LabHorizon gain over the Actor–Simulator–Selector reference is modest—$29.2\%$0 percentage points—even though the improvement over the initial policy is large. The reported immune-science hypothesis has not yet been experimentally tested. Finally, the framework’s acceptance criteria are domain-specific, so the extent to which a common retention principle can be operationalized without substantial manual design remains unresolved.

Conclusion

RSI-Anything formulates recursive self-improvement as a retained policy–world-model system rather than as policy adaptation alone. Its principal mechanism is a closed loop in which predictions prioritize external feedback, observations drive policy or workflow revision and conditional predictor adaptation, and compatibility checks determine which system version persists. Across code repair, GUI interaction, laboratory planning, reaction optimization, and immune-science workflow construction, the reported results show improved feedback allocation, successive policy gains, and higher experimental-search performance under constrained budgets.

The strongest controlled result is the code-repair progression from $29.2\%$1 with a static verifier to $29.2\%$2 with predictive selection and $29.2\%$3 after world-model adaptation, at essentially identical online test usage. The broader results support the framework’s applicability across heterogeneous settings, while leaving open the central systems question of whether policy and world-model adaptation can sustain reliable, jointly improving behavior over long horizons with limited and selectively acquired external feedback.