Test-Time Training
Definition
Test-time training adjusts model parameters based on current inputs and a self-supervised auxiliary objective.
A method that adjusts model parameters during evaluation based on a self-supervised auxiliary objective applied to current inputs.
Simply explained
The model adapts somewhat to new data during evaluation. It uses a training task derived from this data without requiring finished correct answers.
Why it matters in practice
The term helps to select, develop, evaluate AI systems in a well-founded manner and realistically assess their technical risks.
Example from business
An image inspection system adjusts selected parameters to new camera conditions without using production labels.
Limits and pitfalls
- Adaptation can drift, amplify attacks, or influence later cases; it is not merely additional inference effort.
Origin & history
Sun and co-authors introduced Test-Time Training in 2019 as adaptation to distribution shifts.
Use in mid-sized companies
- Possible application: An image inspection system adjusts selected parameters to new camera conditions without using production labels.
Frequently asked questions
What does test-time training mean?
A method that adjusts model parameters during evaluation based on a self-supervised auxiliary objective applied to current inputs. What should be considered? Adaptation can drift, amplify attacks, or influence later cases; it is not merely additional inference effort.
Sources
- Test-Time Training · Sun et al., Test-Time Training
Source checked on 10.10.2026. Editorial application examples are hypothetical; no evidence of real customer results.
Editorially reviewed on: