Self-Generated Feedback Destabilizes Test-Time Training: A Causal Decomposition of Long-Horizon Adaptation

Self-Generated Feedback Destabilizes Test-Time Training. A Causal Decomposition of Long-Horizon Adaptation. By Cheng Luo and two others, from KAUST. Published on the fourth of October. Its code repository is worked out: about forty-five source files, with a readme and a licence.

Test-time training lets a model store information in its weights while it is running inference. The trouble starts when the model learns from its own output. Every update then also changes the model that will generate the next training example, and the loop closes on itself. The claim here is that this self-generated feedback is what destabilizes the process, not the act of writing to the weights.

Across streams of one hundred twenty-eight thousand tokens, keeping updates made on generated text worsens prediction on independent human-written text. That holds for three TTT-E2E model configurations, labeled one hundred twenty-five million, seven hundred sixty million, and three billion, and the same failure appears when Adam updates the existing weights of Qwen3-4B. Yet the same update mechanisms can improve things when the text is real, which rules out writing itself as the cause.

Three matched comparisons trace the causal pathway. Fixed Generation uses a frozen model to generate the training chunks, and removes over ninety-eight percent of the damage at the two smaller sizes. Recorded Replay separates two losses: the one from reading degraded text, and the additional loss that gets stored by updating on it. A paired comparison of a single update then shows the conflict locally. The update predicts its own source better, and new real text worse.

That cost grows after Closed Loop adaptation, and a few trajectories account for most of the large failures. A final check, Settlement, evaluates the candidate state on independent real text before committing it. It leaves mean endpoint gaps of zero point zero seven nats at the smallest size, and minus zero point zero two at the middle one, while keeping the gains from adapting on real text.

The causal story is established for the smaller configurations: the over ninety-eight percent figure and the Settlement gaps are reported for the one hundred twenty-five million and seven hundred sixty million configurations. The streams run to one hundred twenty-eight thousand tokens, and damage is measured on independent human-written text. The failure is specific to learning from generated output. On real text the same mechanisms can help.

The lesson is procedural. An update that looks good on the text that produced it can be bad for everything else, so the authors argue for checking prediction on independent evidence before an update is retained.