Alignment-aware Data Selection for Unlearning in Contrastive Vision-Language Models

Dongjun Hwang, Yejin Kim, Beomyun Kwon, Junsuk Choe

ICML Workshop on MemFM 2026

We observe that unlearning performance in contrastive VLMs largely depends on the composition of the forget set, and propose ALISE—a data selection framework that measures each sample's alignment with both the retain set and the full forget set to facilitate targeted knowledge removal while preserving model utility.

Training-Free Uncertainty-guided Logit Adjustment for Few-Shot Class-Incremental Learning

Sungwon Woo, Dongjun Hwang, Shiwon Kim, Junsuk Choe, Jongho Nang

CVPR Findings 2026

We propose a training-free approach that leverages uncertainty estimates to adaptively adjust logits in few-shot class-incremental learning, effectively preventing catastrophic forgetting without requiring any retraining or modification of the model.

[Paper] [Code]

Enhancing Multi-Image Understanding through Delimiter Token Scaling

Minyoung Lee, Yeji Park, Dongjun Hwang, Yejin Kim, Seong Joon Oh, Junsuk Choe

ICLR 2026

We identify cross-image information leakage as a key bottleneck in Large Vision-Language Models (LVLMs) and propose scaling hidden states of delimiter tokens to reinforce intra-image interaction while suppressing undesired cross-image interference, achieving consistent gains on multi-image benchmarks with no additional training or inference cost.

[Paper] [Code]

OVS Meets Continual Learning: Towards Sustainable Open-Vocabulary Segmentation

Dongjun Hwang, Yejin Kim, Minyoung Lee, Seong Joon Oh, Junsuk Choe

NeurIPS 2025

We address the overlooked problem of continual learning in open-vocabulary segmentation, and propose ConOVS—a Mixture-of-Experts framework that dynamically routes each input to incremental expert decoders, enabling models to recognize new categories without forgetting previously learned ones.

[Paper] [Code]

Does Prior Data Matter? Exploring Joint Training in the Context of Few-Shot Class-Incremental Learning

Shiwon Kim*, Dongjun Hwang*^†, Sungwon Woo*, Rita Singh^†

ICCV Workshop on Continual Learning in Computer Vision 2025

We investigate when and how prior data should be utilized in few-shot class-incremental learning, revealing that naive joint training fails under severe class imbalance, and propose an imbalance-aware joint training benchmark to guide practitioners on optimal training strategies.

[Paper] [Code]

Small Object Matters in Weakly Supervised Object Localization

Dongjun Hwang, Seong Joon Oh, Junsuk Choe

Neurocomputing 2025 SJR Q1; IF: 5.50

We reveal that existing WSOL methods systematically fail to localize small objects, and address this gap by introducing new evaluation metrics and a contrastive learning framework that zooms in on small regions, significantly improving small-object localization without sacrificing accuracy on larger objects.

[Journal] [Code]

Curriculum Learning with Class-label Composition for Weakly Supervised Semantic Segmentation

Dongjun Hwang, Hyoseo Kim, Doyeol Baek, Hyunbin Kim, Inhye Kye, Junsuk Choe

Pattern Recognition Letters 2025 SJR Q1; IF: 3.90

We propose a curriculum learning strategy for weakly supervised semantic segmentation that progressively trains on increasingly complex class-label compositions, effectively alleviating spurious correlations from co-occurring classes and class imbalance across eight WSSS methods on three benchmarks.

[Journal]

Entropy Regularization for Weakly Supervised Object Localization

Dongjun Hwang, Jung-Woo Ha, Hyunjung Shim, Junsuk Choe

Pattern Recognition Letters 2023 SJR Q1; IF: 4.75

We propose a simple entropy regularization term that bridges the gap between classification and localization objectives in WSOL—requiring only one line of code and no architecture changes—consistently improving localization performance across five datasets and multiple backbone networks.

[Journal]