Why Self-Training Helps and Hurts: Denoising vs. Signal Forgetting

A new statistical study reveals why iterative self-training of AI models—where a model is repeatedly trained on its own predictions—creates a fundamental trade-off. The process provides 'denoising' benefits that initially improve performance, but eventually leads to 'signal forgetting' that degrades the model. This creates a U-shaped risk curve, proving there's an optimal early-stopping point. The research provides a data-driven method to find this stopping time, validated on synthetic data.