Google Simula: synthetic data as mechanism design rather than sample-by-sample optimization

Google Research published a detailed technical post on April 16, 2026, about the Simula framework — a synthetic data generation system that fundamentally changes how the problem of data scarcity in specialized AI applications is approached. Authors Tim R. Davidson and Hamza Harkous argue that the problem must be reframed “from the sample level to the mechanism level.”

Why mechanism design rather than sample optimization?

Traditional approaches to synthetic data optimize individual examples — a better prompt, a better temperature, a better filter. The authors argue that this does not scale for domains where data is naturally scarce (regulated fields, novel specialized tasks, privacy-sensitive applications).

Simula instead designs a mechanism that controls the distribution of generated data across multiple axes simultaneously. The result is that practitioners can tune “what the dataset looks like” the way they would design an architecture — with explicit parameters instead of trial and error.

Four control axes

The framework decomposes generation into four independent dimensions:

Global diversification uses reasoning models to build hierarchical taxonomies that map the conceptual space of a domain. These taxonomies serve as “sampling scaffolds” and ensure long-tail coverage instead of clustering around the most common cases.

Local diversification uses meta-prompts derived from taxonomy nodes, generating multiple distinct instances within the same topic to prevent mode collapse — where the model keeps producing variations of the same sample.

Complexification treats difficulty as an orthogonal axis, enabling a shift in the difficulty distribution of the dataset without changing semantic coverage. Practitioners can generate simple and complex variants of the same topic independently.

Quality control operates through a dual-critic loop — two independent verifiers that reduce LLM sycophancy and ensure high-quality labels.

Technical architecture and evaluation

The system uses Gemini 2.5 Flash as the teacher model for generation and Gemma-3 4B as the student model for training. Evaluation relies on Taxonomic Coverage and Calibrated Complexity Scoring metrics, where the latter assigns an Elo rating to each example through LLM batch comparisons.

Tests covered five domains: cybersecurity (CTI-MCQ, CTI-RCM), legal reasoning (LEXam), mathematics (GSM8k), and multilingual knowledge (Global MMLU). Generated datasets had up to 512,000 examples per domain.

An interesting finding: high complexity improves mathematical accuracy by 10% but degrades legal reasoning. The authors interpret this as evidence that “there is no single optimal recipe” — each domain requires its own mix of axes.

Already in production across Google’s ecosystem

Simula is not an experimental project. The post reveals that it already powers:

• Specialized models: ShieldGemma, FunctionGemma, MedGemma
• Safety infrastructure: the primary backbone for Gemini safety classifiers (on-device and server-side)
• User protection: AI fraud detection in Android phone calls and spam filtering in Google Messages
• Enterprise security: frameworks that democratize ML through realistic synthetic attack scenarios

This announcement signals that Google has elevated its internal synthetic data infrastructure to the level of a first-class AI primitive — treating it as seriously as model architecture or hardware stack.