RSI-Anything Ranks Agent Actions Before Paying for Feedback

Recursive self-improvement asks an agent to revise its policy or supporting models across repeated rounds, then use the revised system to generate the next round of behavior. Feedback makes that loop expensive: each candidate may require a hidden test suite, a grader call, an environment interaction, or a physical experiment. The RSI-Anything paper proposes using a predictive world model to rank candidates before execution, reducing the number of calls to the external evaluator.

The framework builds on the group’s earlier JEPA-Anything paper, a domain-agnostic architecture based on Yann LeCun’s joint-embedding predictive approach. A JEPA learns to predict a target’s latent representation from context. JEPA-Anything adds orthogonal predictive factorization, which splits latent targets into complementary factors, learns them through dedicated pathways, and recombines them in a shared predictor. RSI-Anything places an agent-improvement loop around that model.

Feedback is the expensive part

External evaluators limit conventional self-improvement loops in two ways. Graders, tests, simulations, and experiments constrain how many candidates the agent can assess. Meanwhile, policy updates change the agent’s behavior, gradually moving it away from the data distribution used to train its world model. Predictions can become unreliable as that gap grows.

RSI-Anything keeps the external evaluator in the loop while narrowing its workload. The policy generates candidate actions, the world model predicts their outcomes, and the system executes the highest-ranked candidate. The resulting observation supports a policy update and provides a reliability signal for the world model. Component-level and compatibility checks determine whether the updated policy-model pair advances to the next round or rolls back.