Robust defaults for generative optimization across tasks and domains

Establish whether there exist robust, broadly applicable default choices for starting artifacts and learning-context structuring strategies (including which traces to include, truncate, and batch) in large language model–based generative optimization that consistently transfer across different agent designs and application domains to enable broad adoption.

Background

Across case studies, the paper finds that success of generative optimization depends on task-specific choices—starting artifacts, credit horizon, and experience batching—and that no single configuration works universally.

Despite this variability, the authors hypothesize that, akin to widely adopted defaults in traditional machine learning (e.g., Transformers for sequence modeling, Adam for optimization), generative optimization might admit analogous defaults that generalize across tasks, if such defaults can be identified.