The Evolutionary Designer: LLM-Driven Multitask AVOs for Cross-Domain Optimization

TL;DR: Imagine an AVO agent that can evolve solutions across wildly different domains (like scheduling, design, and robotics), using multitask learning and generative models. By fusing LLM-driven generative multitasking (like LLM2TEA) with the AVO paradigm, we could enable cross-domain transfer, creativity, and rapid adaptation.

Research Question: Can multitask AVO agents, leveraging LLM-based generative representations, facilitate transfer and innovation across heterogeneous optimization domains?

Hypothesis: Multitask AVOs equipped with transferable representations and generative capabilities will outperform single-domain agents in both solution diversity and adaptability, especially when exposed to domain shifts or composite objectives.

Experiment Plan: Construct an AVO multitask agent drawing on LLM2TEA (Wong et al., 2024) for cross-domain genotype/phenotype generation and fitness evaluation. Evaluate in at least two distinct optimization domains (e.g., 3D design and logistics scheduling), with transfer tasks and composite objectives. Compare diversity, novelty, and performance against domain-specific AVOs and classical multitask evolutionary algorithms.

References:

• Chen, T., Ye, Z., Xu, B., Ye, Z., Liu, T., Hassani, A., et al. (2026). AVO: Agentic Variation Operators for Autonomous Evolutionary Search.
• Wong, M., Liu, J., Rios, T., Menzel, S., & Ong, Y. (2024). LLM2TEA: An Agentic AI Designer for Discovery With Generative Evolutionary Multitasking. IEEE Computational Intelligence Magazine.