Epidemiology of Model Collapse: Modeling Synthetic Data Contamination via Bilayer SIR Dynamics

Core summary

One sentence to understand this update

Research models synthetic data contamination leading to model collapse using bilayer SIR dynamics, revealing cross-contamination issues in the AI ecosystem.

Impact & opportunity

What this could mean

LLM trainers and data managers should be aware of the long-term impact of synthetic data on model performance and explore mitigation strategies to prevent model degradation.