Microsoft Research team introduces Improvement Dynamics Model, boosting self-evolving agent performance by up to 46%

Researchers from Microsoft Research, the University of Virginia, and Tsinghua University have introduced a new method called the Improvement Dynamics Model (IDM) to optimize the code modification process for self-evolving agents. This approach allows agents to predict the impact of proposed modifications on specific tasks before undergoing expensive evaluations, enabling optional refinements based on these forecasts.

Traditional self-evolving agents typically rely on trial and error, receiving feedback only after completing time-consuming and costly evaluations. The IDM accumulates predictive knowledge by leveraging past modifications and their observed outcomes, thereby improving forecast accuracy across generations. The research team tested this method using the HyperAgents framework across four domains: multilingual coding, paper review, robotics reward design, and Olympiad-level mathematical grading.

The results show that agents equipped with the IDM achieved an average performance improvement of up to 46% compared to the baseline HyperAgents across all four domains, maintaining this advantage even under matched token budgets that include both evolution and evaluation costs. Additionally, the learned predictive knowledge transferred to improve evolution in new domains. This study is a preprint and has not yet been independently reproduced.