POM: Pretrained Optimization Model wins zero-shot black-box tasks

Xiaobin Li and colleagues from Xidian University (with co-authors from other institutions) introduced Pretrained Optimization Model (POM), a neural network-based optimizer designed for zero-shot black-box optimization. Traditional optimizers such as evolutionary strategies require careful hyperparameter tuning when applied to new problems, and often struggle to generalize to unseen task distributions. POM sidesteps this by pretraining on a large corpus of diverse optimization tasks, learning a general strategy that can be directly applied to new tasks—or quickly adapted via few-shot fine-tuning.

In experiments on the BBOB benchmark and two robot control tasks, POM outperformed state-of-the-art black-box optimization methods, particularly on high-dimensional problems where prior approaches tend to degrade. The paper reports that fine-tuning POM with only a modest number of samples and a limited budget yields substantial performance improvements. Additionally, POM showed robust generalization across varied task distributions, dimensions, population sizes, and optimization horizons. The authors have released their code, making it straightforward for practitioners to integrate POM into their own optimization pipelines and evaluate its zero-shot performance on custom problems.