Synthetic Data Needs an Evidence Contract

TL;DR for operators

Synthetic data should have a defined job before anyone scales its production. For model training, the relevant test is whether generated examples add nonredundant learning signal and improve held-out performance without unacceptable regressions. For consumer research, the test changes: statistically diverse text is not enough if the business claim concerns what real customers believe.

Two recent studies reach this conclusion from different directions. Better-structured synthetic training data can continue improving model knowledge after simpler recipes plateau. But synthetic consumer responses can achieve human-like diversity statistics while still differing in language, topic emphasis, and representational meaning.

The operating rule is therefore not “generate more” or even “maximize diversity.” Specify what the synthetic data must preserve, engineer generation around that requirement, and validate against the truth standard of the downstream decision.

Synthetic data is often discussed as though it were a raw material: produce enough examples, use a stronger generator, remove obvious defects, and quality should rise with volume. That assumption is attractive because generation is measurable and inexpensive relative to collecting many forms of real data.

The difficulty is that generated volume can increase while useful information barely changes. A corpus may repeat the same training signal in different wording. A synthetic panel may produce varied prose without becoming more representative of the people it is meant to describe.

For businesses, those failures lead to different costs. A domain-model team can spend substantial compute training on redundant supervision. A market-research team can make a more serious interpretive error by treating plausible synthetic opinions as evidence about actual customers. Both problems begin when generation quality is confused with evidentiary validity.

Training data has to add learning signal

Han and colleagues test synthetic data as a mechanism for placing domain knowledge into model parameters.¹ Their baseline result is a useful warning: several existing synthetic question-answer and document-generation approaches show diminishing returns as synthetic-token volume grows. Using a much larger generator also does not reliably remove the plateau.

The alternative is to change the structure of the generated data.

The paper combines synthetic question-answer examples with synthetic documents during continued training. The proposed explanation is that the two formats push learning in different directions. Question-answer examples appear to teach both domain information and more transferable behavior for recalling and using knowledge, while document-style training provides additional exposure to domain facts. Gradient comparisons support this complementarity, although the authors do not claim a complete mechanistic decomposition.

That distinction matters because mixing several document-rewriting methods produced little additional gain. The useful variation was not simply more stylistic forms of the same material.

The authors also introduce Focal Rewriting, in which document rewrites are conditioned on particular questions. This increased measured lexical and semantic diversity and produced a steeper scaling curve. Here, diversity has a concrete function: it reduces repeated topical coverage as generation volume increases.

Across the reported experiments, the combined recipe beat vanilla retrieval-augmented generation, or RAG, in five of six model-benchmark settings, with an average 2.6% relative accuracy advantage. RAG supplies relevant external documents when the model answers a query; the synthetic-training approach instead tries to internalize more of the domain knowledge.

The more operational result is that the two approaches remained complementary. Adding RAG to the synthetically trained models improved all six reported settings, averaging 5.9 percentage points and a 9.1% relative gain over vanilla RAG.

This evidence does not establish that retrieval can be removed from production architectures. It supports a different allocation decision: for products repeatedly working over relatively stable proprietary knowledge, some domain familiarity can be trained into the model while retrieval remains available for source context and additional accuracy.

Diversity is not itself a validity test

France and Albinsson examine synthetic data in a setting where the required truth is fundamentally different: consumer research.² They compare LLM-generated projective responses with human responses from 173 college students evaluating five U.S. tourism destinations.

The generation process is highly configurable. The study varies model, sampling temperature, prompting strategy, projective task, and human few-shot examples. Those controls materially change the resulting synthetic corpus.

Higher temperature generally increases vocabulary and lexical diversity while reducing concentration. Expanded prompts also increase the reported diversity measures. Human seed examples move synthetic responses closer to the human benchmark in style and content. Model choice produces substantial differences in verbosity, vocabulary, and diversity.

A team measuring only aggregate diversity could therefore tune the system until the numbers look reassuring.

The topic analysis shows why that is insufficient. Some synthetic conditions reproduced many broad associations found in human responses, yet their emphasis still differed. The models also produced stylized compound descriptors that increased apparent lexical variety without necessarily adding an equivalent amount of substantive consumer information.

In selected settings, synthetic responses could approach or even exceed human responses on particular diversity metrics. That did not make them equivalent evidence about consumers.

This is the central distinction between distributional similarity and representational validity. Two corpora can have similar entropy, lexical concentration, or topic overlap while still encoding different meanings or being generated through different processes. Most importantly, matching a narrow human sample does not establish that synthetic respondents represent a wider customer population.

The paper therefore positions synthetic consumer responses as support for ideation, supplementation, and research piloting. It explicitly does not justify using them to estimate how prevalent a belief is among real consumers.

Give each synthetic dataset an evidence contract

The two papers can be combined into a practical framework because the relevant validation target changes with the downstream job.

Cognaptus inference: teams should write this contract before selecting the generator or setting a synthetic-token budget.

For a domain assistant, the contract might specify that generated training data must improve a fixed suite of domain questions, preserve general capabilities within defined regression tolerances, and be compared with a retrieval-only deployment. Generation settings that fail those tests are low-value even if they produce millions of fluent examples.

For a consumer-insight workflow, the contract should specify a different endpoint. Synthetic respondents can expand the hypothesis space before expensive fieldwork, but any statement such as “customers associate the product with X” or “30% of buyers believe Y” requires evidence from an appropriate human sample. Synthetic generation can reduce the cost of discovering what to investigate; it does not automatically supply the population being investigated.

The budget should include design and validation

This framing also changes how synthetic-data projects should be budgeted.

In the training study, larger target models needed fewer synthetic tokens to reach the reported RAG-level performance threshold. For Qwen3, fitted curves estimated requirements of roughly 102 times the original corpus token count for the 14B model, compared with 142 times for 8B, 177 times for 4B, and 813 times for 1.7B.

That creates a resource-allocation problem. A smaller deployment model may be cheaper per inference but require considerably more synthetic-data generation and continued-training effort to reach a target knowledge level. Generator cost, target-model capacity, training compute, retrieval cost, and serving cost should therefore be evaluated together.

The consumer study creates a parallel validation expense. Changing the model, prompt, temperature, or seed examples changes the resulting “consumer” evidence. A production research workflow therefore needs version tracking and periodic calibration against human benchmarks. Treating synthetic respondents as a static dataset property ignores the configurable system that produced them.

Neither paper eliminates the need for real data. The training study uses source documents as the grounding material and leaves catastrophic forgetting only partially examined. The consumer study retains human responses as the reference and explicitly limits population inference.

What changes is where scarce real information is used. Synthetic generation can multiply and reorganize available signal. The scarce evidence should then be reserved for the places where the system cannot validate itself: general-capability regressions, factual grounding, human prevalence, lived experience, and other claims whose truth exists outside the generator.

Build for the claim you need to make

Synthetic data can reduce information-production costs, but only after the downstream claim is specified.

If the goal is to make a model answer domain questions better, generated data can be judged by whether training on it produces reliable gains. Complementary supervision, topical coverage, model capacity, and retrieval all become engineering variables.

If the goal is to understand people, model-generated responses can help enumerate possibilities and test research designs. They cannot establish how widespread those possibilities are without human evidence that supports that inference.

The distinction is operational, not philosophical. Before scaling synthetic generation, write down the user of the output, the decision it will affect, the property the data must preserve, and the external test that can falsify an inadequate dataset.

Once those are explicit, synthetic-data quality becomes measurable. Without them, higher volume and higher diversity are simply properties of the generator.

Cognaptus: Automate the Present, Incubate the Future.