Self-Generated Feedback Destabilizes Test-Time Training
October 3, 2026
Research shows that when Test-Time Training (TTT) models learn from their own generated output, prediction accuracy on human-written text degrades. This causal decomposition demonstrates that using a frozen model for generating training chunks removes over 98% of this damage.
HOW THIS AFFECTS YOU
●
builderIf implementing TTT, use a frozen model to generate training data to avoid model degradation.
●
researcherYou must account for the feedback loop instability when designing long-horizon TTT models.