Simula: A Smarter Way to Build Synthetic Datasets from Scratch

We’ve all been there. You need a specialized dataset for some niche AI application, and the well of public data is bone dry. Manual annotation is soul-crushing, expensive, and error-prone. Synthetic data has been the promised savior, but most approaches feel like duct-taping prompts together and hoping for the best. Google Research’s new paper, “Reasoning-Driven Synthetic Data Generation and Evaluation,” published in TMLR, takes a different angle. They’ve built a framework called Simula that treats dataset creation like mechanism design, not just sample generation.

The core problem with existing synthetic data methods is they operate at the sample level. You generate one data point, then another, maybe tweak a few parameters, but you never really design the dataset as a coherent whole. Coverage is an afterthought. Complexity is a byproduct. And if you need edge cases? Good luck. Simula flips this by working from first principles: it uses reasoning models to map out the entire conceptual space of a domain before generating a single example.

Here’s where it gets interesting. Instead of relying on human-written seeds or evolutionary black boxes, Simula builds deep, hierarchical taxonomies recursively. The system proposes candidate sub-categories, then a critic model evaluates, merges, and filters them. This propose-and-refine loop runs until you have a dense tree of concepts covering the long tail of your domain. Think cyber threat intelligence, medical anomalies, or rare manufacturing defects. The taxonomy becomes a sampling scaffold, giving you control over global diversity without clustering around the common modes.

Once you’ve got that scaffold, you can dial in complexity and quality independently. That’s the part I like. Most synthetic data generators entangle these variables, so making data harder also makes it noisier. Simula separates them: you can crank up reasoning complexity without sacrificing label quality, or generate simple examples with high precision. This is a massive win for stress-testing models. Instead of waiting for failures to happen in production, you proactively generate edge cases from the taxonomy’s tails.

The paper walks through several experiments, and the results are solid. Simula outperforms baseline synthetic data methods across multiple benchmarks, especially in domains where data is scarce or privacy-sensitive. The generated datasets are more diverse, more structured, and more useful for fine-tuning specialized models. The reasoning-first approach also means that as underlying models get smarter, Simula’s generation capabilities improve naturally. No retooling, no new prompts.

Is it perfect? No. The taxonomy generation step is computationally heavy, and the critic model can introduce its own biases if not carefully tuned. But for anyone building AI for niche applications, this feels like a step in the right direction. We need tools that treat data as code — versioned, reproducible, and inspectable. Simula gets closer to that ideal than anything I’ve seen recently.