Test-Time Training

Given new data at deployment, taking a few self-supervised update steps on it before predicting with the updated model.

Test-time training was proposed by Sun and colleagues at ICML 2020: the model is trained with an auxiliary self-supervised task alongside its main one, for instance predicting how much an image was rotated, and at test time, given a new sample, it first takes a few gradient steps on that auxiliary task before making its real prediction, to cope with a mismatch between training and test distributions. A related idea, test-time adaptation, such as Tent (ICLR 2021), instead just minimizes prediction entropy on the test data and adjusts only the normalization layers' parameters. Embodied AI uses this to let a policy keep adapting on-site after deployment: EVOLVE-VLA uses automatically estimated task progress as feedback to keep a VLA learning at test time, and TTT-Parkour first scans and reconstructs unfamiliar terrain, then quickly fine-tunes a humanoid parkour policy on the reconstruction.

ExampleTTT-Parkour (2026) uses an RGB-D camera to scan and reconstruct unfamiliar obstacles such as wedges and narrow beams, then fine-tunes a humanoid robot's parkour policy on the reconstructed terrain; the paper reports most terrain takes under 10 minutes from capture through reconstruction to test-time training.

Also called

TTT, Test-Time Adaptation, TTA

Related

Out-of-Distribution · Domain Adaptation · Self-Supervised Learning · Inference-Time Compute · Continual Learning · Rapid Motor Adaptation

Sources

Sun et al. 2020: Test-Time Training with Self-Supervision for Generalization under Distribution Shifts (ICML 2020)
Bai, Gao, Shou 2025: EVOLVE-VLA: Test-Time Training from Environment Feedback for Vision-Language-Action Models
Zhu et al. 2026: TTT-Parkour: Rapid Test-Time Training for Perceptive Robot Parkour

As of

2026-02