Google’s Simula: Building Synthetic Datasets Like You’re Designing a Market

Google Research just dropped a paper that makes me feel a little better about the state of synthetic data. It’s called Simula, and instead of the usual “throw a bigger model at it” approach, they’re treating dataset generation like designing an economic mechanism. That’s way more interesting than it sounds.

The real problem with synthetic data

We all know the drill. Generalist AI models feast on internet-scale data, but specialized applications? Healthcare, finance, safety-critical systems? Those domains are data deserts. Manual annotation is expensive and slow. Real-world data is brittle—once you collect it, you’re stuck with whatever distribution you got, including all its biases and blind spots.

Most synthetic data pipelines today are glorified prompt factories. You seed them with examples, run some evolutionary algorithm, and pray the distribution matches your need. But that’s sample-level thinking. You’re optimizing one data point at a time, not designing the whole dataset. It’s like building a house by stacking bricks without a blueprint.

Mechanism design for data

Simula flips the script. It treats dataset generation as a mechanism design problem—you define what you want (coverage, complexity, quality) as separate levers, then let a reasoning model figure out how to pull them. No seeds. No black-box evolution. Just first-principles reasoning.

The key insight is that you can’t just ask for “more data.” You need control. Simula gives you four knobs:

• Global diversification: Instead of random sampling, it builds a deep taxonomy of the domain first. A reasoning model recursively proposes sub-categories, a critic model filters and merges them. The result is a hierarchical map that covers the long tail, not just the popular clusters.

• Local complexity: Once you know what categories you need, you can dial the difficulty per sample. Want easy examples for onboarding and hard ones for stress-testing? You set the complexity budget, and Simula generates accordingly.

• Quality constraints: This is where the mechanism design part shines. You can impose constraints like “no personally identifiable information” or “all examples must be internally consistent.” The model optimizes for these alongside coverage and complexity.

• Scalability without seeds: Because the whole process is reasoning-driven, it scales with model capability. Better reasoning models produce better taxonomies and better samples. No human-in-the-loop required once you define the constraints.

Why this matters

I’ve seen too many synthetic data projects fail because they optimized for quantity over structure. You end up with a million samples that all look the same, or a thousand that are technically correct but useless for training. Simula’s taxonomy-first approach is a genuine improvement.

It also addresses a pet peeve of mine: the reactive safety culture in AI. We wait for models to fail, then patch them. Synthetic data lets you generate edge cases before they happen—adversarial inputs, rare events, corner cases. Simula makes that proactive generation systematic rather than ad-hoc.

The catch

It’s not magic. The framework relies on the underlying reasoning model’s capability. If your model can’t build a good taxonomy, your dataset won’t be good either. And the mechanism design formulation adds complexity—you need to define your constraints clearly, which isn’t always easy.

But compared to the alternatives (manual curation, seed-dependent generation, evolutionary algorithms that no one can explain), this is a step forward. It’s programmable data. Version it. Reproduce it. Inspect it. That’s what production workflows need.

The paper is in TMLR, and the framework is called Simula. If you’re building in data-scarce domains, it’s worth a read. I’m planning to test it on a medical imaging dataset next week—let’s see if the theory holds up in practice.