Test-time adaptation is a family of methods that adjust a model during evaluation or deployment instead of leaving it completely fixed after training. The adjustment may use the current inputs, feedback or repeated attempts to improve performance on a new distribution or task.
The method can make benchmark progress difficult to attribute to scale alone. A result may reflect the model, the adaptation procedure, additional computation and the way researchers structured the test, so reports should describe the full experimental setup.
ELI5
Test-time adaptation changes how an AI model behaves or learns while it is being evaluated or used. The model can adjust to the current inputs, feedback, or repeated attempts instead of staying completely fixed after training.
For example, a model may update part of its behavior after seeing several examples from a new data source. Reported performance should identify the extra adaptation and computation rather than attributing every gain to the base model.
How is test-time adaptation different from training?
Conventional training happens before deployment on a prepared dataset. Test-time adaptation makes limited adjustments while the model encounters evaluation or real-world inputs, often without rebuilding the model from scratch.
Why can test-time adaptation improve benchmark results?
It can help a model adjust to the structure or distribution of the current task. The improvement belongs to the complete method, including adaptation and computation, not necessarily to the base model alone.

