What Drives Test-Time Adaptation for CLIP? A Controlled Empirical Study from a Update Perspective

Jiazhen Huang^†, Xiao Chen^†, Zhiming Liu^†, Yaru Sun, Jingyan Jiang, Zhi Wang

We show that adaptation gains primarily arise from test-time evidence and reliable proxies rather than heavy optimization, and the most effective adaptation paradigm varies with the type of distribution shift.

^† indicates equal contribution (co-first authors).

Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction

Jiazhen Huang, Zhiming Liu, Changhu Wang, Wei Ju, Ziyue Qiao, Xiao Luo

We identify the brittleness of transductive methods under imbalanced distributions and propose MOON, a training-free, model-agnostic framework that dynamically adjusts shrinkage strength to mitigate negative transfer.

Test-Time Distillation for Continual Model Adaptation

Xiao Chen^†, Jiazhen Huang^†, Zhiming Liu, Qinting Jiang, Fanding Huang, Jingyan Jiang, Zhi Wang

We propose a collaborative test-time distillation framework for continual model adaptation that improves robustness and generalization under realistic distribution shifts.

^† indicates equal contribution (co-first authors).

Adaptive Disentangled Representation Learning for Incomplete Multi-View Multi-Label Classification

Quanjiang Li^†, Zhiming Liu^†, Tianxiang Xu^†, Tingjin Luo, Chenping Hou

We proposed ADRL, a novel framework that jointly addresses structural distortion and semantic ambiguity in incomplete multi-view settings by integrating label-guided feature disentanglement and category-aware embedding interaction.

^† indicates equal contribution (co-first authors).