Self feedback loops undermine test-time training in language models
Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation
Computation and LanguageArtificial Intelligence
Summary
When a language model learns from its own writing during use, changes it makes can cause it to perform worse on real human text later. The authors show that this happens because the model’s updates affect the very text it uses to train itself, creating a feedback loop that degrades performance over time. They test different setups and find ways to reduce the problem by checking new model states on independent, real text before keeping updates. This suggests it’s important to verify improvements on fresh evidence before letting a model adapt.
What this means in practice
- •For language model engineers: Develop safeguards to prevent performance drops during model updates on self-generated text.
- •For machine learning operations teams: Implement validation steps evaluating model updates on independent real data before deployment.
Authors
Cheng Luo, Bing Li, Bernard Ghanem
Abstract
Test-time training (TTT) lets a model store information in its weights during inference. When the model learns from its own output, however, each update also changes the model that generates the next training example. Across 128K-token streams, retaining generated-text updates worsens prediction on independent human-written text with three TTT-E2E model configurations (labeled 125M, 760M, and 3B). The same failure occurs when Adam updates Qwen3-4B's existing weights. The same update mechanisms can improve on real text, so writing itself is not the failure. Three matched comparisons trace the causal pathway. Fixed Generation removes over 98% of the damage at 125M and 760M by using a frozen model to generate training chunks. Recorded Replay separates the loss caused by reading degraded text from the additional loss stored by updating on it. A paired one-update comparison then shows the local conflict: an update predicts its source better but new real text worse. This cost grows after Closed Loop adaptation, with a few trajectories accounting for most large failures. Finally, Settlement evaluates the candidate state on independent real text before commitment. It leaves mean endpoint gaps of 0.07 and -0.02 nats at 125M and 760M while retaining real-text adaptation. These results motivate checking prediction on independent evidence before retaining an update.