# Model collapse from synthetic data

Area: [Synthetic data and self-improvement for LLMs](https://mlsanity.com/synth/index.md). HTML: https://mlsanity.com/synth/model-collapse-from-synthetic-data/

## Key facts

- Model collapse from synthetic data: 85 papers on arXiv since 2023 (data through 7 Oct 2026).
- 28 of its 85 mapped papers (33%) cite “The Curse of Recursion: Training on Generated Data Makes Models Forget” (2023, arXiv 2305.17493).
- 26 of its 85 mapped papers (31%) cite “A Tale of Tails: Model Collapse as a Change of Scaling Laws” (2024, arXiv 2402.07043).
- 7 new papers 17 Sep–7 Oct.

## Most cited here

- [The Curse of Recursion: Training on Generated Data Makes Models Forget](https://mlsanity.com/paper/2305.17493/) (2023), cited by 28 papers here
- [A Tale of Tails: Model Collapse as a Change of Scaling Laws](https://mlsanity.com/paper/2402.07043/) (2024), cited by 26 papers here
- [Model Collapse Demystified: The Case of Regression](https://mlsanity.com/paper/2402.07712/) (2024), cited by 22 papers here
- [Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data](https://mlsanity.com/paper/2404.01413/) (2024), cited by 22 papers here
- [How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse](https://mlsanity.com/paper/2404.05090/) (2024), cited by 18 papers here
- [Strong Model Collapse](https://mlsanity.com/paper/2410.04840/) (2024), cited by 17 papers here
- [Beyond Model Collapse: Scaling Up with Synthesized Data Requires Verification](https://mlsanity.com/paper/2406.07515/) (2024), cited by 13 papers here
- [Position: Model Collapse Does Not Mean What You Think](https://mlsanity.com/paper/2503.03150/) (2025), cited by 10 papers here
- [Rate of Model Collapse in Recursive Training](https://mlsanity.com/paper/2412.17646/) (2024), cited by 6 papers here
- [Recursive Learning Without Collapse: A Weighting-Based Stabilization Framework](https://mlsanity.com/paper/2502.18049/) (2025), cited by 5 papers here

## New (17 Sep–7 Oct)

- [From Behavior to Provenance: Attributing Tabular Foundation Models to Synthetic Pretraining Data](https://mlsanity.com/paper/2610.02347/) (2026-10-01)
- [No Model Required: Text Entropy Rate Filtering Mitigates Iterative Fine-Tuning Collapse](https://mlsanity.com/paper/2610.01493/) (2026-10-01)
- [Feature Selective Model Collapse in Diffusion Models: Total Replacement versus Fixed-Budget Training](https://mlsanity.com/paper/2610.01318/) (2026-10-01)
- [Effective Synthetic Data Curation Requires Group-Level Signals](https://mlsanity.com/paper/2610.00779/) (2026-09-30)
- [Source Identification Is Not Fitness Testing: Measuring the Limits of Synthetic-Data Attribution](https://mlsanity.com/paper/2610.00417/) (2026-09-30)
- [Shared Autoregressive Context Can Distort Relationships in Synthetic Data](https://mlsanity.com/paper/2609.32546/) (2026-09-26)
- [Let Training Guide Selection: Online Synthetic Data Filtering via Real-Anchored Utility](https://mlsanity.com/paper/2609.29988/) (2026-09-24)