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

AI summary

Overview

• Research area: Machine learning systems for recursive self-improvement (RSI), predictive world models, and domain-general agent frameworks spanning code generation, GUI interaction, laboratory planning, chemistry, and immune science.
• Technical level: Advanced. The paper formalizes a full problem specification (task tuples, selection rules, adaptation triggers, acceptance gates) and uses latent-predictive modeling notation, though the high-level mechanism is describable without the math.
• Scope: One sentence — the paper proposes a framework that couples prediction-guided action selection with feedback-driven, gated retention of policy–world-model pairs, and reports gains across five applied settings.

What This Paper Is About

Recursive self-improvement systems need external feedback (test runs, experiments, evaluations) to get better, but that feedback is expensive, and a world model's predictions can drift out of date as the policy that generates candidates keeps changing. RSI-Anything addresses this by using a predictive world model to rank which candidates are worth spending feedback on, then using the resulting observations to update either the policy, the world model, or both — while component-level and pair-level checks decide whether the updated system is actually retained. The goal is a domain-general loop in which each accepted system revision shapes the next round of interaction and learning.

Key Contributions

1. World models inside recursive self-improvement. The framework, built on the JEPA-Anything predictive foundation, uses world-model predictions to anticipate action outcomes and guide which experience gets collected. Improvement is formulated at the level of the policy–world-model system rather than either component alone, making prediction reliability under evolving policies a first-class design requirement.
2. A recursive update-and-retention mechanism. Prediction-guided action selection and learning-data selection are connected to feedback-driven policy learning or workflow revision, plus conditional world-model adaptation. Component and compatibility checks decide which policy–model pair is retained, allowing the two components to evolve asynchronously.
3. A reliability-triggered adaptation protocol with evidence categories. Adaptation fires only when reliability assessment returns unreliable (cases of insufficient coverage do not trigger it), when adaptation data is sufficient, and when budget allows. Evidence is explicitly typed: E0 deterministic rule checks, E1 learned-model predictions and rollouts, E2 specialized simulators or higher-fidelity calculations, and E3 external measurements or experiments.
4. Empirical evidence across five settings. Gains are reported for code repair, GUI task completion, laboratory planning across five successive policy versions, reaction-yield optimization at matched experiment counts, and an end-to-end immune-science application.

Main Findings

• Code repair improves in two successive steps. With the ReasonFlux-Coder-7B policy held fixed and four candidate revisions proposed before one is executed, final hidden-test success on 200 CodeContests tasks (three seeds, up to three feedback rounds) rose from 26.3% with a capacity-matched static verifier to 29.2% (175/600) with world-model-guided selection, then to 31.2% after world-model adaptation. Random selection scored 26.0%. The paired gain over the static verifier was +2.83 percentage points (95% interval [+0.17, +5.50]).
• Better prediction accompanies better repair. In the fixed-world-model versus static-verifier comparison, candidate-ranking accuracy rose from 66.5% to 72.8%, and the five-class Brier score fell from 0.488 to 0.358. Online test-suite usage was comparable: 3.71 calls per task for the fixed world model (CURE-WM-Lite), 3.72 for the static verifier, and 3.72 for the adaptive configuration.
• Gains generalize to other code benchmarks. JEPA-Anything-guided candidate selection reached 46.37% final full-test success on APPS Introductory (1,000 problems) and 33.05% on LiveCodeBench v6 (1,055 problems), exceeding random selection and the static verifier under the same coder and candidate budget.
• GUI task completion rises while action counts fall. On MiniWoB++ click-checkboxes, success went from 51% at baseline to 79% (adapted) to 99% (final). Mean actions per episode dropped from 5.30 to 3.21 (a 39.4% reduction), repeated clicks fell from 237 to 0, and episodes with incorrect submissions fell from 9 to 1. The adapted configuration used fewer actions (2.93 per episode) but produced 21 incorrect-submission episodes. The World-based control scored 49%, 96%, and 98% across the three configurations.
• OSWorld gains are relative percentages. RSI-Anything improved task success rate by 5.5% relative to RLAnything and by 18.4% relative to the initial agent; RLAnything itself improved success by 12.3% relative to the initial agent.
• Laboratory planning improves monotonically across five retained versions. On the same 200-task LabHorizon F200 Level-2 benchmark, Final Scores for five consecutive versions of a 4B policy from one self-improvement process were 21.80%, 38.03%, 41.99%, 44.56%, and 46.01% — a 24.21 percentage-point gain from first to last. The final version also exceeded the Actor–Simulator–Selector reference (45.32%) by 0.69 percentage points, and reached 0.700 Level-1 accuracy versus 0.665 for the reference.
• Reference-model context (LabHorizon). Table 3 lists Grok 4.3 (0.555 L1, 0.3244 L2), Gemini 3.1 Pro (0.465, 0.3263), GPT-5.5 (0.535, 0.2276), Kimi K2.6 (0.550, 0.3150), Qwen3.6-35B-A3B direct (0.475, 0.2534) and trained (0.635, 0.4100), and RSI-Anything 4B (0.700, 0.4601).
• Reaction yields beat Bayesian optimization at matched experiment counts. In the GC-analyzed Pd–NHC–DMS Suzuki–Miyaura screen, best observed yield after 20 experiments was 93.0% for RSI-Anything versus 87.5% for Bayesian optimization (gain of 5.5 percentage points). In the flow LC–MS screen using Pd(OAc)₂, mean best observed yield after 40 experiments was 96.15% versus 82.16% (gain of 13.99 percentage points).
• An end-to-end immune-science pipeline produced a concrete hypothesis. Agent Genesis searched scientific-agent workflows and selected a model-building workflow that produced an action-conditioned immune world model representing cellular, tissue, and individual immune states. During analysis, self-correction removed the screened cytokine-pair candidates, and the integrated evidence led to a prospective nomination combining SIRPα inhibition with IL-36γ-axis modulation.

Methodology in Plain English

The system starts from JEPA-Anything, a cross-domain predictive foundation that uses joint-embedding predictive architectures with orthogonal predictive factorization (OPF): latent targets are split into complementary factors, predicted separately from context, and recombined into a complete latent state.

RSI-Anything then wraps a loop around that predictor:

1. Propose. The current policy generates a set of candidate actions, patches, or workflows.
2. Predict and rank. The world model predicts outcomes for each candidate. Valid candidates are scored by a weighted utility rule, optionally minus an uncertainty penalty (or, if enabled, an explicitly information-seeking acquisition rule can favor uncertain candidates instead). Invalid candidates are filtered out before ranking.
3. Evaluate one. Only the selected candidate is actually executed or evaluated, producing a paired prediction–observation record. This is what keeps external feedback cheap.
4. Update. Feedback drives a policy update (demonstration learning, preference learning, or workflow search) or a world-model adaptation — but adaptation only fires when the reliability check on predictions attributable to the current model version returns unreliable, the adaptation data is sufficient, and the remaining budget covers the estimated cost. The two components can update asynchronously.
5. Accept or reject. A component check compares each proposed component against its current version. If at least one component changed, a pair check compares the assembled candidate pair against the current pair under matched evaluation conditions; a failed pair check retains the entire current pair. If neither component changed, the pair carries forward without re-evaluation.
6. Retain and repeat. The retained pair shapes the next round's candidates and learning experience.

Separately, imagined trajectories can be admitted into simulated training data only if they pass validity, minimum-utility, uncertainty, and permitted-evidence-source filters — and this training filter does not block external evaluation of uncertain candidates. Evaluation data, gate data, and a held-out audit set are kept distinct.

Why This Matters

The paper's central claim is that improving a policy and improving its world model independently does not guarantee a better joint system, because each shifts the other's operating conditions. Treating the pair as the unit of retention, and gating adaptation on measured reliability, is the distinctive mechanism.

Research impact: It reframes recursive self-improvement as a systems problem involving reliability, coverage, and acceptance criteria rather than pure policy optimization, and it supplies a formal update-and-retention protocol that other work can adopt or critique.

Real-world applications:

• Code repair assistants that decide which of several candidate patches is worth spending a test-suite run on.
• GUI automation agents that complete interface tasks with fewer actions and fewer incorrect submissions.
• Laboratory and protocol planning, where action sequences must satisfy reagent, instrument, and ordering constraints.
• Chemistry and drug discovery, where each experiment is expensive and candidate conditions must be ranked before running them.

Industry relevance: Any organization running agents against a metered budget — cloud test runners, robotics, laboratory automation, clinical or biological experimentation — faces the same tension between feedback cost and improvement quality. A framework that adapts its predictor only when predictions are demonstrably unreliable, and that rolls back the whole pair when compatibility fails, is directly relevant to deployment safety and cost control. The immune-science case shows the loop operating end to end on a scientific discovery workflow rather than a benchmark.

Future Directions

The paper does not report a dedicated future-work section in the available content, but the design raises clear open questions:

1. What happens when reliability assessment returns insufficient? The protocol explicitly does not trigger adaptation in that case, so how the system should safely explore to widen coverage remains unresolved.
2. How far do the gains extend beyond three feedback rounds and five policy versions? The code-repair evaluation caps at three rounds and the laboratory-planning study at five retained versions, leaving long-horizon recursion untested in the reported results.
3. How transferable is the retained pair across domains? Each domain implementation specifies its own action encodings, outcome readouts, and enabled pathways; whether shared predictors transfer between them is not reported.
4. Can the nominated immune-science hypothesis be validated? The SIRPα plus IL-36γ-axis nomination is described as prospective — the paper reports a hypothesis for experimental testing, not a confirmed therapeutic result.

Target Audience

Researchers and engineers working on self-improving agents, world models, and model-based planning will find the framework formalization and the update-and-retention protocol most useful. Practitioners building agents under constrained feedback budgets — code-generation tooling, GUI automation, lab automation, or reaction-condition optimization — benefit from the matched-budget comparisons against static verifiers and Bayesian optimization. Readers interested in the scientific-discovery angle will find the immune-science application the most concrete case study. The paper assumes familiarity with latent predictive architectures, so beginners should expect to consult the JEPA-Anything foundation first.

Authors’ 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.