Daily AI Paper

Recursive self-improvement is a deceptively simple idea: a system does a task, learns from the experience, and then becomes better at the next round of doing and learning. The promise is obvious. If the system can keep turning feedback into lasting upgrades, then it may improve not just within a task, but across tasks and across time. The hard part is that real feedback is expensive. You usually cannot test every candidate action, every possible workflow, or every proposed solution in the real world. And once the policy changes, the system’s behavior shifts, which can make any predictor or world model less reliable. So the central challenge is not just “can we learn from feedback,” but “can we keep improving recursively when feedback is limited and the system itself is changing?” This paper tackles that challenge with a framework called RSI-Anything. The key idea is to combine recursive self-improvement with predictive world models. A world model is a predictor of what will happen if the system takes a certain action in a certain context. In broad terms, that lets the system imagine consequences before paying the full cost of acting. In AI research, this is a familiar idea in planning and model-based reinforcement learning, but here the paper pushes it into a more general setting: not just one domain, but a framework that can be instantiated across code repair, GUI interaction, laboratory planning, reaction optimization, and even an immune-science workflow. To understand why this matters, it helps to separate two kinds of improvement. One is immediate improvement within a task: you try candidate actions, observe outcomes, and use that feedback to choose better actions next time. The other is persistent improvement across rounds: the policy, workflow, or agent itself gets updated so that future tasks start from a better place. Recursive self-improvement needs both. But there is a coupling problem. If you update the policy, the distribution of actions changes, so the world model may no longer predict well. If you update the world model, the ranking of candidates can change, which changes what feedback you collect and how the policy learns. So the system has to decide when to update the policy, when to update the model, and when to keep a given policy-model pair rather than accept a new one. RSI-Anything builds on an earlier predictive framework called JEPA-Anything. The important background here is joint-embedding predictive architectures, which learn to predict target representations from context rather than directly predicting every raw detail. JEPA-Anything adds an orthogonal predictive factorization idea, where the target representation is decomposed into complementary factors that can be predicted separately and then recombined. That gives a flexible world-model foundation that can work across heterogeneous domains. RSI-Anything inherits that predictive backbone, but adds the recursive improvement loop around it. The framework has three main pieces. First, the current policy or workflow proposes candidate actions. Second, the world model predicts the consequences of those candidates, or in some settings predicts a bounded rollout over several steps. Third, external feedback is used to update either the policy, the world model, or both. The paper emphasizes that these updates are not automatic or unconditional. There are component-level checks that ask whether a policy update or model update is individually acceptable, and there are compatibility checks that ask whether the updated policy-model pair should actually be retained for the next round. That retained pair is important, because it determines what candidates get proposed next and what experience is gathered next. In other words, the system does not just learn from history; it learns which version of itself should generate the next history. A nice feature of the framework is that it distinguishes different kinds of evidence. Some checks are deterministic rule checks. Some come from the learned world model itself. Some use stronger simulators or higher-fidelity computation. And some use real external measurements or experiments. This matters because not every piece of evidence should be treated equally. The framework also separates two uses of feedback: one set of records may be used to improve the policy, while another may be used to adapt the predictor. Those sets can overlap, but they do not have to be identical. On the selection side, the world model helps decide which candidates are worth spending expensive feedback on. If a policy proposes several candidates, the predictor estimates their likely outcomes, and a verifier checks validity and constraints. Then the system ranks feasible candidates using predicted utility, optionally penalized by uncertainty. In some settings, the framework can also prioritize uncertain candidates if the goal is information gathering rather than direct utility. This is a practical point: when feedback is expensive, the best use of the budget is not always to test the candidate that looks best. Sometimes it is better to test the candidate that most reduces uncertainty or most improves the learning signal. The experimental story is organized around several different domains, which is important because the paper is making a domain-general claim, not just a narrow benchmark claim. In code repair, the system ranks multiple proposed revisions before executing one. In GUI interaction, the agent has to complete tasks through action sequences, so the framework evaluates whether predictive selection can improve task completion while reducing unnecessary actions. In laboratory planning, the paper looks at successive policy versions across five rounds, testing whether retained updates really compound over time. In reaction optimization, the framework is used in a scientific setting where the objective is to find high-yield experimental conditions under a fixed experiment budget. And in an immune-science application, the paper connects workflow evolution, predictive modeling, and therapeutic hypothesis generation end to end. The results are presented as evidence that world-model-guided selection and recursive retention do in fact help. In code repair, with the coder fixed, the paper reports that success rises from 26.3% with a static verifier to 29.2% when predictive selection is added, and then to 31.2% after world-model adaptation, at comparable test usage. The important takeaway is not just the absolute numbers, but the pattern: using predictions to choose what to evaluate is better than a static verifier alone, and adapting the world model further improves performance. That supports the core claim that prediction and update need to work together, rather than treating the model as a fixed auxiliary component. In GUI evaluations, the framework achieves higher task completion and reduces average actions by 39.4%. That second result is especially meaningful because it suggests the agent is not merely acting more, but acting more efficiently. In interactive environments, fewer actions for the same or better success rate usually means the system is making better decisions about what to try and when to stop. The laboratory-planning setting is the clearest demonstration of recursive gains across versions. Over five successive policy versions, planning scores increase from 21.80% to 46.01%. That more than doubling is notable because it indicates that retained updates are compounding rather than resetting the system each round. In other words, the framework is not just finding a one-off good policy; it is preserving improvements in a way that benefits later versions. The reaction-optimization result is also strong. Under the same 40-experiment budget, RSI-Anything achieves a mean best observed yield of 96.15%, compared with 82.16% for Bayesian optimization. Here the message is that predictive selection and recursive update can be highly competitive in a classic experimental optimization setting, where every experiment is expensive and sample efficiency matters a lot. The comparison to Bayesian optimization is useful because it situates the framework against a well-established method for black-box optimization under limited trials. The immune-science application is less about a single scalar metric and more about demonstrating integration. The paper says it connects workflow evolution and predictive modeling with therapeutic hypothesis generation. That matters because scientific workflows are often multi-stage and iterative: analyze data, revise the workflow, update the model of the domain, and generate new hypotheses. RSI-Anything is designed to support exactly that sort of loop. There are, of course, important limitations and things to watch next. The biggest conceptual limitation is that the framework depends on prediction reliability under evolving policies. If the policy changes too much, the world model may drift out of date. The paper addresses this with conditional adaptation and compatibility checks, but that does not eliminate the underlying challenge. Another limitation is that the framework is only as good as the domain-specific instantiation: action encodings, verifiers, readouts, budgets, and update routes all have to be defined carefully for each application. So although the framework is domain-general, it is not domain-free. A second point is that the reported gains depend on the available budgets and the specific evaluation setup. For example, the code repair result is reported at comparable test usage, and the reaction optimization comparison is under a fixed 40-experiment budget. That is the right way to evaluate sample efficiency, but it also means the conclusions should be read in the context of those budgets and tasks. Still, the broader contribution is clear. RSI-Anything shows that recursive self-improvement can be framed as a coupled system of prediction, selection, feedback, and retention. The world model is not just a simulator; it is a guide for where to spend scarce feedback. The policy is not just a generator; it is something that can be revised and retained. And the combination of the two, when managed with compatibility checks, can produce measurable gains across very different kinds of problems. That makes the paper interesting not just as a collection of benchmarks, but as a concrete step toward systems that improve themselves in a controlled, persistent, and increasingly general way.