A frozen AI model can imitate an unseen dataset from one example

Given a single reference sequence from a domain it was never trained on, a pretrained generator produces new synthetic samples that match its statistics and temporal structure, with no retraining.

Chinese Tech

Yujie Li · Zezhi Shao · Jingbo Zhou · Yongjun Xu · Xueqi Cheng · Fei Wang

State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Baidu Frontier Research Department

Research Digest·3 November 2026·3 min read

Illustration: A frozen AI model can imitate an unseen dataset from one example

In one line

A pretrained generator can create realistic new time series for an unseen domain from a single reference sequence, without retraining, by reusing learned shape and shape-transition vocabularies.

Before you read this

Two ideas this paper builds on

synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.64 papers →self-supervised learningA paradigm where models learn representations from unlabeled data by predicting hidden or transformed parts of the input.51 papers →

Read it as

Generating believable fake time series, such as sensor readings or electricity demand, usually needs many real examples from the same setting. When those examples are scarce or private, it fails.

MarkovTS is trained in advance on many kinds of data, learning a small vocabulary of recurring local shapes in a signal and of which shapes tend to follow which. Given one reference sequence from an unseen setting, it reuses that vocabulary to produce new sequences with the same rhythm but fresh randomness, without retraining.

On 12 unseen datasets, its samples matched the references better than strong alternatives in most comparisons, and adding them as extra training data improved forecasting, classification and anomaly detection.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

flow matching synthetic data generation self-supervised learning