Graph Models: The Lean, Mean Machines in Human Simulation

Large language models (LLMs) like GPT-3 have made waves in simulating human behavior. But do we really need them for every task? A fresh contender, graph neural networks, might just be the lean, mean machine we've been waiting for.

The GEMS Approach

Enter Graph-basEd Models for Human Simulation (GEMS). This approach views closed-ended simulation as link prediction on a heterogeneous graph of individuals and choices. While this sounds technical, the reality is simple: GEMS connects the dots between people and their decisions without relying on bulky language models.

Here's what the benchmarks actually show: GEMS matches or even surpasses top LLMs across three datasets and evaluation settings. And it does so using three orders of magnitude fewer parameters. That's not just efficient, that's revolutionary. In an age where computational resources are stretched thin, this could be a major shift.

Why This Matters

The numbers tell a different story about what's possible beyond large language models. For tasks like survey predictions and test-taking, GEMS offers a transparent and efficient alternative. The architecture matters more than the parameter count. If GEMS can perform as well as big models without the resource drain, why aren't we using it more?

Think about the implications. Less data means less energy consumption, which is essential as we become more conscious of AI's environmental impact. With its efficient architecture, GEMS isn't just a technological advancement. it's an ecological one too.

Looking Ahead

Where do we go from here? Should we continue pouring resources into developing even larger LLMs, or do we pivot towards smarter, leaner solutions like GEMS? The choice seems clear. The future of human simulation might not lie in scaling up but in scaling smart. Frankly, it's time to strip away the marketing hype and focus on what works best.