Predict Before You Modify: Learning Improvement Dynamics for Self-Evolving Agents

A predictive self-evolution approach uses an editable Improvement Dynamics Model (IDM) to forecast task-level outcomes of proposed modifications before evaluation, guiding optional revisions and accumulating predictive knowledge across generations. Evaluated on multilingual coding, paper review, robotics reward design, and Olympiad-level mathematical grading, the method outperforms HyperAgents by up to 46% average performance over generations.