Kibok Lee

Initializing prompts in flatter regions of the loss landscape dramatically improves calibration and performance in test-time prompt tuning for vision-language models.

VLMs learn faster and better when you dynamically weight the prefixes based on input token importance, rather than treating all tokens equally.

Stop throwing away the baby with the bathwater: selectively resetting model parameters during test-time adaptation prevents catastrophic forgetting and collapse, leading to better long-term performance.