Synthetic Data in the Era of Large Language Models

Vijay Viswanathan, Xiang Yue, Alisa Liu, Yizhong Wang, Graham Neubig2025 · DOI

Abstract

Progress in natural language processing (NLP) has historically been driven by the availability of high-quality data. Today, researchers are increasingly turning to "synthetic data"—data generated with the assistance of large language models (LLMs)—to accelerate dataset construction and reduce costs. However, much of this work remains ad hoc, with researchers often "reinventing the wheel" rather than building on established foundations. This paper outlines a systematic framework for synthetic data generation, categorizing recent progress into four key pillars: generation algorithms, general-purpose development, scenario-specific applications, and critical analysis of open limitations.

Introduction: The Shift to Synthetic Data

The evolution of NLP is inextricably linked to data. Traditionally, high-quality datasets were the product of labor-intensive, human-centric processes. While effective, these methods are often too slow and expensive to keep pace with the rapid development of LLMs. The pivot toward synthetic data offers a promising alternative, yet the field currently lacks a shared methodology. Without a unified framework, it is difficult to compare results, ensure data quality, or understand the long-term implications of training models on machine-generated content.

A Unified Framework for Synthetic Data

To move beyond fragmented, one-off generation scripts, we propose a pedagogical framework organized around four central pillars.

1. Generation Algorithms

The first pillar focuses on the technical methods used to produce synthetic data. The goal is to move beyond simple prompting to sophisticated algorithms that ensure the resulting data is high-quality, diverse, and accurate. This involves developing techniques that can filter noise and verify the factual integrity of generated samples.

2. General-Purpose Development

Synthetic data is not just a substitute for human data; it is a tool for enhancing the core capabilities of language models. This pillar explores how synthetic datasets can be used to improve instruction following, reasoning, and general-purpose performance. By using LLMs to generate complex reasoning chains, researchers can target specific model weaknesses that are difficult to address with standard web-scraped data.

3. Scenario-Specific Customization

While general capabilities are important, many applications require niche or domain-specific data. This pillar examines how synthetic data generation can be tailored for specialized tasks, allowing for the creation of high-quality training sets in domains where human expertise is scarce or expensive to procure.

4. Critical Analysis and Open Questions

The final pillar addresses the risks and limitations inherent in synthetic data. Key concerns include:

  • Model Collapse: The risk that models trained on synthetic data may eventually lose diversity or converge on errors.
  • Bias Amplification: The potential for synthetic data to reinforce or exacerbate biases present in the generating model.
  • Foundational Rigor: The need for standardized metrics and evaluation protocols to ensure the reliability of synthetic datasets.

Key Insights: From Ad Hoc to Systematic

The transition from "clever prompts" to a rigorous engineering discipline is the primary shift required in the field. This involves synthesizing insights from NLP and related disciplines to build a shared understanding of data-centric reasoning. By focusing on the foundations of data production, the community can improve the effectiveness of synthetic data, moving beyond simple text imitation toward the development of robust, reasoning-capable models.

Conclusion

Synthetic data represents a frontier in LLM research, but its potential can only be fully realized through systematic study and foundational rigor. By unifying recent advances into a structured framework, researchers can better address the challenges of quality, bias, and long-term model health, ensuring that synthetic data serves as a reliable building block for the next generation of AI.

Summary updated

Abstract

Progress in natural language processing has historically been driven by better data, and researchers today are increasingly using "synthetic data" -data generated with the assistance of large language models -to make dataset construction faster and cheaper.However, most synthetic data generation approaches are executed in an ad hoc manner and "reinvent the wheel" rather than build on prior foundations.This tutorial seeks to build a shared understanding of recent progress in synthetic data generation from NLP and related fields by grouping and describing major methods, applications, and open problems.Our tutorial will be divided into four main sections.First, we will describe algorithms for producing high-quality synthetic data.Second, we will describe how synthetic data can be used to advance the general-purpose development and study of language models.Third, we will demonstrate how to customize synthetic data generation to support scenario-specific applications.Finally, we will discuss open questions about the production and use of synthetic data that must be answered to overcome some of their current limitations.Our goal is that by unifying recent advances in this emerging research direction, we can build foundations upon which the community can improve the rigor, understanding, and effectiveness of synthetic data moving forward.
BibTeX
@misc{viswanathan2025synthetic,
  author = {Vijay Viswanathan and Xiang Yue and Alisa Liu and Yizhong Wang and Graham Neubig},
  title = {{Synthetic Data in the Era of Large Language Models}},
  year = {2025},
  doi = {10.18653/v1/2025.acl-tutorials.7},
}
Figure