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.

References

Viewed through this lens, we conjecture that with sustained research explorations generative optimization may eventually admit robust ``defaults'' that enable broad adoption: just as Transformers~\citep{vaswani2017attention} provided a broadly useful inductive bias for sequence modeling, we may discover starting artifacts for agents that are broadly optimizable across tasks; and just as Adam~\citep{kingma2014adam} works well across a wide range of neural architectures, we may discover robust ways to structure the learning context --- what traces to include, truncate, and batch --- that transfer across agent designs and domains.

— Understanding the Challenges in Iterative Generative Optimization with LLMs  (2603.23994 - Nie et al., 25 Mar 2026) in Conclusion and Discussion

The agentic studiers and frozen solver use models from the same family and share the Claude Code harness, so we cannot establish whether the resulting artifacts transfer across solvers or provide solver-independent improvements.

— Studying Without a Syllabus: Task-Agnostic Environment Preprocessing  (2609.10824 - Samuel et al., 9 Sep 2026) in Discussion and Limitations, Section 6

Because the ranker that provides the reward signal shares a model family with the components optimized against it, the learned behaviors may partly capture ranker-specific preferences rather than ranker-agnostic optimization knowledge. We probe this with a cross-ranker evaluation in Section~\ref{sec:cross_ranker}; fully disentangling transferable knowledge from ranker-specific artifacts remains future work.

— GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search  (2610.23709 - Kim et al., 16 Oct 2026) in Limitations section