Can predictive world models enable recursive self-improvement across diverse domains?

Prediction alone does not close the improvement loop

Recursive self-improvement (RSI) means turning interaction experience into changes that persist and shape the next round of action and learning. The paper's starting point is that a world model can predict the consequences of candidate actions, but a prediction by itself does not improve anything. Predicted outcomes must guide which candidates receive costly external feedback, and that feedback must produce updates that are retained.

The difficulty is that the two components move together. A policy update changes which actions and experience the world model sees, so its predictions may become less reliable. A world-model update changes candidate rankings and imagined trajectories, which in turn changes the learning signal for the policy. Improving either component alone therefore does not guarantee better joint behavior, which is why the framework treats the policy-model pair, not each part separately, as the object that is accepted or rejected.

Predictions decide which candidates get tested

RSI-Anything builds on JEPA-Anything, a shared world-model construction framework that predicts target representations from observed context. In the instantiated system, the current policy proposes a set of candidates, and the world model supplies an outcome prediction or a bounded rollout for each. A selection rule then ranks feasible candidates by predicted utility, optionally penalizing uncertainty, and returns a candidate or batch whose evaluation cost fits the remaining feedback budget.

The code-repair experiment isolates this step. The coder is held fixed, so any change comes from how revisions are chosen rather than from a better policy. The policy proposes 20 revisions, and the instantiated rule aggregates predicted execution-class probabilities to rank four proposed revisions before one is executed. This is the concrete meaning of predictive selection here: the world model does not replace the test, it decides where the limited test budget goes.

Selection and adaptation add separate gains in code repair

With the coder fixed, success rises from 26.3% with a static verifier to 29.2% through predictive selection, and then to 31.2% after world-model adaptation, at comparable test usage. The ordering matters: the first gain comes from choosing which revision to execute, and the second comes from adapting the predictor to observed outcomes. Because the coder never changes, neither number should be read as evidence that the policy itself improved.

Adaptation is conditional rather than automatic. The framework assesses prediction reliability against feedback history using the current model version, and only attempts adaptation when the assessment is unreliable, the adaptation data are sufficient, and the remaining budget covers the estimated cost. Insufficient coverage yields an insufficient assessment rather than evidence of reliability, and does not trigger adaptation. This design keeps a predictor from being retrained on the basis of feedback that does not actually cover the candidates being considered.

GUI tasks complete more while taking fewer actions

In graphical user interface evaluations, RSI-Anything updated both the policy and the world model and showed higher task completion, while average actions decreased by 39.4%. The discussion adds that on the checkbox task the system had fewer repeated clicks and incorrect-submission episodes. These are different measures from the code-repair success rate: task completion and action count describe behavior in an interactive environment, not the fraction of repaired programs.

The GUI setting is also the one where both components change, so it is the closest of the reported evaluations to the full loop. Even so, the reported comparison is between the updated system and its alternatives on these tasks, not a demonstration that policy and world model kept improving each other over many rounds within a single task.

Retained policy versions more than double lab-planning scores

Laboratory planning follows five successive policy versions, and the score rises from 21.80% to 46.01% across them. This is the clearest evidence in the paper for the retention part of the argument: improvement accumulates because accepted policy versions carry forward and shape later decisions, rather than because a single update was large.

The scope is stated by the authors themselves. The laboratory results assess protocol-aligned planning, and reliable physical execution would require evaluation on the relevant equipment and workflows. The doubling therefore describes planning scores under the paper's protocol, not demonstrated performance of laboratory work carried out by the system.

A fixed experiment budget favors predictive search

In reaction optimization, RSI-Anything reaches 96.15% mean best observed yield, compared with 82.16% for Bayesian optimization under the same 40-experiment budget. The shared budget is what makes the comparison informative: both methods are judged on the best yield found after the same number of experiments, so the difference reflects how the search used those experiments rather than how many were run.

This result belongs to feedback-guided condition search, a different task from code repair and GUI control. It supports the claim that prediction can prioritize which experiments to run when measurements are expensive, but it does not by itself show that the policy and world model improved each other across rounds in that chemistry setting.

The joint loop remains partly untested

The evaluations instantiate distinct parts of the protocol rather than the whole protocol at once. Code repair holds the coder fixed while testing predictive selection and predictor adaptation; laboratory planning follows successive policy versions; GUI evaluations update both components; chemistry tests feedback-guided condition search; and the immune-science application separates workflow evolution and model construction from downstream analysis with a fixed predictor.

The authors accordingly note that joint multiround evaluation is needed to establish sustained mutual improvement of the policy and world model within a single task. That is the main boundary on the headline question: the paper shows that prediction-guided selection, conditional model adaptation, and retained policy versions each produce measured gains in their own settings, not that a single system compounded both kinds of improvement over many rounds.

Prioritize selection when feedback is the scarce resource

The most decision-relevant comparison in this paper is the code-repair sequence, because it holds the policy fixed and varies only how feedback is spent. Predictive selection alone accounts for the move from 26.3% to 29.2%, and adaptation adds a further 2.0 percentage points to 31.2% at comparable test usage. For a reader deciding where to invest, that ordering suggests the selection step is the first thing to get right when external evaluations are the binding constraint.

The reaction-optimization result points the same way under a hard budget: 96.15% versus 82.16% mean best observed yield at 40 experiments. The condition attached to both findings is the one the framework itself enforces. Adaptation should be attempted only when reliability checks on the current model version show unreliable predictions, sufficient adaptation data, and available budget; otherwise the retained pair should stay as it is.

Sources used in this answer