This is a fascinating project! It's exciting to see the application of multi-agent simulations to a real-world problem like optimizing marketing messages. Your approach of grounding the AI personas in public data and focusing on social influence is very insightful.

I recently authored a paper on a related topic that might be of interest, particularly concerning your challenges with accuracy and long-term simulation stability. Our work, "SALM: A Multi-Agent Framework for Language Model-Driven Social Network Simulation," introduces a novel framework for integrating language models into social network simulations.

I specifically focused on achieving long-term temporal stability in multi-agent scenarios.[2] A couple of our key contributions could be relevant to your work:

I developed a hierarchical prompting architecture that enables stable simulations beyond 4,000 timesteps while significantly reducing token usage.[3]

To address memory growth and maintain agent consistency, I implemented an attention-based memory system that achieves high cache hit rates with sub-linear memory growth.

I also established formal bounds on personality stability, which could be a useful concept for ensuring the accuracy and predictability of your AI personas over time.

Our validation against SNAP ego networks demonstrated the capability of our framework to model long-term social phenomena with empirically validated behavioral fidelity.[3]

It seems like there's a lot of synergy between our research and your product. I'd be very interested to hear your thoughts on our paper and explore if any of our findings could be beneficial for the continued development of Artificial Societies.

Here's the link to my paper: https://arxiv.org/abs/2505.09081