Demystifying Synthetic Data in LLM Pre-training: A Systematic Study of Scaling Laws, Benefits, and Pitfalls (2025.emnlp-main)
Copied to clipboard
Feiyang Kang, Newsha Ardalani, Michael Kuchnik, Youssef Emad, Mostafa Elhoushi, Shubhabrata Sengupta, Shang-Wen Li, Ramya Raghavendra, Ruoxi Jia, Carole-Jean Wu
| Challenge: | a large-scale empirical study compares natural web data, diverse synthetic types, and mixtures of natural and synthetic data. |
| Approach: | They conduct a large-scale empirical study on large-volume LLMs using a unified protocol and scaling laws. |
| Outcome: | The proposed method is faster than pre-training on natural web data, the authors show . their results are consistent with previous studies on rephrased text and textbooks . |
Similar Papers
Synthetic Data Generation with Large Language Models for Text Classification: Potential and Limitations (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have explored using large language models to generate synthetic datasets . however, the effectiveness of the LLM-generated synthetic data is inconsistent across different classification tasks. |
| Approach: | They propose to use large language models to generate synthetic datasets to better understand factors that moderate the effectiveness of LLM-generated synthetic data. |
| Outcome: | The results show that subjectivity is negatively associated with the performance of the model trained on synthetic data. |
Scaling Data-Constrained Language Models with Synthetic Data (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large language models (LLMs) improve with more training data, but practical limitations on data collection constrain further scaling. |
| Approach: | They compare three strategies to generate Japanese text, repeat the limited Japanese Web text, and use English Web text to fill the data shortfall. |
| Outcome: | The proposed model outperforms baselines and achieves the performance achieved when the entire token budget is filled with additional organic Japanese Web text. |
On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey (2024.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) provide a data-centric solution to alleviate limitations of real-world data with synthetic data generation. |
| Approach: | They propose a generic workflow for LLM-driven synthetic data generation. |
| Outcome: | The proposed workflows highlight gaps in existing research and outline avenues for future studies. |
Better as Generators Than Classifiers: Leveraging LLMs and Synthetic Data for Low-Resource Multilingual Classification (2026.findings-eacl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable multilingual capabilities, making them promising tools in both high- and low-resource languages. |
| Approach: | They use a multilingual LLM to generate synthetic datasets covering 11 languages and 4 classification tasks and use them to train smaller models. |
| Outcome: | The proposed model outperforms the large generator in low-resource languages and tasks. |
Scaling Low-Resource MT via Synthetic Data Generation with LLMs (2025.emnlp-main)
Copied to clipboard
Ona de Gibert, Joseph Attieh, Teemu Vahtola, Mikko Aulamo, Zihao Li, Raúl Vázquez, Tiancheng Hu, Jörg Tiedemann
| Challenge: | a recent study has shown that LLM-generated synthetic data can improve low-resource machine translation performance . traditional data augmentation techniques like back-translation preserve the human-written target and synthesize the other . |
| Approach: | They construct a document-level synthetic corpus from English Europarl and extend it via pivoting to 147 additional language pairs. |
| Outcome: | The proposed model can significantly improve low-resource machine translation performance even when noisy. |
Understanding the Influence of Synthetic Data for Text Embedders (2025.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in general purpose text embedders have been driven by training on synthetic training data. |
| Approach: | They propose to use GPT-4 to produce high quality synthetic data that expands existing training datasets for embeddings to new tasks. |
| Outcome: | The proposed dataset is high quality and leads to consistent improvements in performance. |
A Survey on Efficient Large Language Model Training: From Data-centric Perspectives (2025.acl-long)
Copied to clipboard
Junyu Luo, Bohan Wu, Xiao Luo, Zhiping Xiao, Yiqiao Jin, Rong-Cheng Tu, Nan Yin, Yifan Wang, Jingyang Yuan, Wei Ju, Ming Zhang
| Challenge: | achieving data-efficient post-training of Large Language Models is a key research question. |
| Approach: | They propose a taxonomy of data-efficient LLM post-training methods from a data-centric perspective. |
| Outcome: | The proposed methods cover data selection, data quality enhancement, synthetic data generation, data distillation and compression, and self-evolving data ecosystems. |
How Much Pretraining Does Structured Data Need? (2026.eacl-long)
Copied to clipboard
| Challenge: | Large language models are increasingly adopted for handling structured data, despite pretraining on unstructured text. |
| Approach: | They propose to re-initialize subsets of layers with random weights before fine-tuning on structured datasets. |
| Outcome: | The proposed models are compared to unstructured datasets and show that they perform well over structured data. |
Demystifying Mixed Outcomes of Self-Training: Pre-training Analyses on Non-Toy LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Recent studies on self-training report seemingly contradictory outcomes. |
| Approach: | They use OLMo-2 models as non-toy LLMs and perform multiple rounds of continual pre-training using self-generated text with different prompting strategies and data filtering. |
| Outcome: | The proposed model collapse is inherent to the training procedure itself, while self-improvement is likely owes its success to human-designed, strategic synthetic pipelines that inject external intelligence. |
Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling (2024.acl-long)
Copied to clipboard
| Challenge: | Large language model pre-training is infeasible due to the large compute costs and duration associated with pre- training and the impending scarcity of high-quality data on the web. |
| Approach: | They propose to use an off-the-shelf instruction-tuned model prompted to paraphrase documents on the web in specific styles such as “like Wikipedia” or in “question-answer format” to jointly pre-train LLMs on real and synthetic rephrases. |
| Outcome: | The proposed model speeds up pre-training by 3x on the C4 dataset, and improves perplexity by 50% on average across different subsets of the Pile. |