Synthetic Data

Synthetic data refers to data artificially generated through algorithms rather than collected directly from the real world; it simulates the statistical characteristics of real data distributions but does not correspond to real individuals or events.

When language models are repeatedly trained on data generated by themselves or similar models, systematic degradation occurs, with the diversity of model outputs decreasing with each generation and the tail information of the distribution gradually being lost.

If the synthetic data generation process omits correlations between key variables, models trained on such data often perform poorly in real-world deployment environments.

合成数据在长尾场景补充和隐私敏感场景替代两个维度上价值最高，而非作为主体数据源

A common approach for generating synthetic data is to have a model rewrite, expand, or role-shift based on real user conversations, producing training samples that are fictitious in form but retain the semantic information of the original.

Research published in 2023, such as "The Curse of Recursion," shows that when models are repeatedly trained on their own generated data, the data distribution gradually degrades, diversity continuously decreases, and model capabilities systematically decline (model collapse).

Procedural synthetic data can generate on demand edge-case scenarios that are difficult to collect in reality—such as arbitrary-intensity glare, low illumination, and heavy occlusion—and provides precise annotations.

Procedural synthetic data provides robotics and autonomous driving developers with a low-cost, high-precision, and infinitely scalable data acquisition path

The vast majority of teams do not record the provenance of each sample (teacher model, prompt template, filters) when generating synthetic data, causing data curation to become blind guessing and potentially entrenching the teacher's blind spots.

A ten-million-row synthetic dataset with no lineage is strictly worse than a hundred-thousand-row dataset in which every sample is traceable, because only the latter is debuggable.

When reviewing samples that mix real and synthetic text, human annotators anchor on the surface fluency of synthetic text, thereby systematically favoring synthetic samples.

Same-family filters (such as a reward model trained on teacher outputs) systematically favor fluent, well-formatted samples, resulting in variance reduction rather than bias correction.

In its pretraining report, Microsoft emphasized the role of data curation and synthetic data in breaking through the data wall.

Generative models cannot produce information beyond their learned parameters; every synthetic generation introduces sampling variance and truncates rare tokens, causing downstream training to accumulate statistical bias, and the recursive loop is essentially a process of information loss.

Logic42's consulting practice has observed over the past 12 months multiple enterprise teams using frontier models to generate synthetic domain training corpora in order to cut annotation budgets, with severe cognitive degradation appearing within three generations of fine-tuning.

In synthetic data, replacing identifiers, names, section order, and surface wording can create samples that appear different but retain the same underlying phrase templates, which provides shortcuts for the model; exact duplicate detection cannot find this overlap, and finer-grained methods such as n-gram overlap, embedding similarity, or template family tracking are needed.

Synthetic data lacks complexities such as real camera noise, motion blur, and non-Lambertian surfaces, so a model's advantages on it may not directly extrapolate to real captured scenes.

The test tasks in this study are described as synthetic data, not derived from real user scenarios.

The current work is only a proof of concept, and all results are based on synthetic simulation data and have not yet been tested for generalization across plasma regions or on real experimental measurements.

Common means of avoiding synthetic data blind spots include: sampling anonymized production data and feeding it back, explicitly injecting same-name/same-title samples into the test set, doing property-based testing on the deduplication logic, and tracking business-level metrics such as 'number of unique titles returned per query' in CI.