Agentic-TTT: Training test-time policy for test-time training
Jiahao Lu and Mohan Kankanhalli (NUS AI Institute, National University of Singapore) present Agentic-TTT, which trains a policy to decide for each query whether to run test-time training, which TTT procedure to call, and whether to reuse a previously built parameter-space skill.
Ask this paper
TTT as tools. Heterogeneous test-time training procedures are exposed as callable tools, and the parameter updates they produce are stored as reusable skills that become part of the deployment environment.
No fixed method works everywhere. Each TTT algorithm helps in some domains and hurts in others, so applying one procedure to every query wastes compute or lowers accuracy.
Results. On the authors' eight-domain benchmark, Agentic-TTT raises Qwen2.5-7B utility from 0.256 with direct inference to 0.510, with gains in every domain. Guidance distillation alone reaches 0.375 and outcome-driven RL adds the rest.
Compute control. Training with compute penalties from 0 to 0.4 produces policies that adapt more selectively, trading utility against TTT cost.
Transfer. The policy generalizes to data sources not seen in training and across model families and scales.
Abstract
Test-time training (TTT) adapts an LLM's parameters using signals derived from test inputs, and can make striking improvements in pre-specified settings such as IMO competitions or designated open problems. By turning deployment experience into parameter updates, TTT provides a direct mechanism for model-level self-improvement. Yet TTT is not universally beneficial: each TTT algorithm works in different settings, and applying an ill-suited method could waste test-time compute or even damage model performance. Therefore, such parameter-level self-improvement requires agency: the model must decide when TTT is warranted, which algorithm to invoke, and whether an existing skill can be reused. To fill this gap, we introduce Agentic-TTT, which learns a test-time policy to govern those decisions. Agentic-TTT turns TTT procedures into callable tools, treats accumulated skills as an evolving deployment environment, and trains its policy using the observed utility gains from its decisions. On our benchmark, Agentic-TTT nearly doubles the utility over the backbone model, learns to trade off utility against compute, and generalizes to domains unseen during training. Together, these results point toward autonomous self-improvement: models that can decide how to learn from their own deployment experience.