Selected work

Representative Papers

Unsupervised Long-Tailed Adaptation of Vision-Language Models

Oct 06, 2026

This study addresses the severe performance degradation on head classes in vision-language models (VLMs) during unsupervised long-tailed adaptation caused by distribution mismatch. For the first time, this work formally defines this task and elucidates its underlying impairment mechanisms. To overcome these challenges, we propose MARS, which innovatively leverages a zero-shot VLM as a fixed reference. By integrating margin-preserving alignment with perception-gap self-refinement strategies, MARS effectively mitigates pseudo-label bias, corrects inherent model biases, and refines predictions for tail and easily confused classes. Extensive experiments across nine benchmark datasets demonstrate that MARS achieves an average accuracy improvement of 4.71%, significantly outperforming existing state-of-the-art methods.

0 citationsRead paper

Hub for Outliers, Spokes for Inliers: Uniform Latent Space Construction for Dual-Mismatched Semi-Supervised Learning

Oct 05, 2026

This study addresses the dual-mismatch problem in semi-supervised learning caused by imbalanced class distributions and the intrusion of unknown classes within unlabeled data. We propose a unified framework integrating a hub-spoke geometric structure with an evidential classifier. Specifically, the latent space is organized into a hub-spoke topology: known-class features are uniformly distributed across peripheral nodes to enhance discriminability, while unknown-class samples are aggregated at the central hub based on low evidential support for effective isolation. This structured feature organization suppresses majority-class dominance bias and substantially improves pseudo-label quality. Extensive experiments demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches across various settings, achieving performance gains of up to 3.25%.

0 citationsRead paper

Recent publications

Latest Papers

Unsupervised Long-Tailed Adaptation of Vision-Language Models

Oct 06, 2026

This study addresses the severe performance degradation on head classes in vision-language models (VLMs) during unsupervised long-tailed adaptation caused by distribution mismatch. For the first time, this work formally defines this task and elucidates its underlying impairment mechanisms. To overcome these challenges, we propose MARS, which innovatively leverages a zero-shot VLM as a fixed reference. By integrating margin-preserving alignment with perception-gap self-refinement strategies, MARS effectively mitigates pseudo-label bias, corrects inherent model biases, and refines predictions for tail and easily confused classes. Extensive experiments across nine benchmark datasets demonstrate that MARS achieves an average accuracy improvement of 4.71%, significantly outperforming existing state-of-the-art methods.

0 citationsRead paper

Hub for Outliers, Spokes for Inliers: Uniform Latent Space Construction for Dual-Mismatched Semi-Supervised Learning

Oct 05, 2026

This study addresses the dual-mismatch problem in semi-supervised learning caused by imbalanced class distributions and the intrusion of unknown classes within unlabeled data. We propose a unified framework integrating a hub-spoke geometric structure with an evidential classifier. Specifically, the latent space is organized into a hub-spoke topology: known-class features are uniformly distributed across peripheral nodes to enhance discriminability, while unknown-class samples are aggregated at the central hub based on low evidential support for effective isolation. This structured feature organization suppresses majority-class dominance bias and substantially improves pseudo-label quality. Extensive experiments demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches across various settings, achieving performance gains of up to 3.25%.

0 citationsRead paper