AI-assisted MemComputing: Scientific Research as Meta-Optimization (Part 1)

Scientific research can be viewed as a hierarchy of optimization problems: well-defined studies maximize explicit performance metrics such as accuracy or efficiency, whereas exploratory studies seek new patterns or physical insights. We formalize this hierarchy as meta-optimization—optimizing what to optimize. An inner loop improves a system’s performance for a fixed metric, while an outer loop searches for the metric itself by maximizing an alignment objective that measures how well a candidate metric predicts, generalizes, and causally tracks the underlying scientific goal.

We implement this concept with a multi-agent large language model system that autonomously plans, codes, and analyzes experiments under a Monte Carlo graph-search controller. Each stage becomes an explicit optimization task connected through learned influence weights. Using MemComputing as a testbed, we show that this framework can automatically design new algorithms and discover emergent performance metrics that reveal deeper physical structure. (Part 2 will discuss implementation details and results.)