Zhihui Zhu

Understanding Task Vectors in In-Context Learning: Emergence, Functionality, and Limitations

Learning to Adapt: In-Context Learning Beyond Stationarity

From Compression to Expression: A Layerwise Analysis of In-Context Learning

AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Features

From Emergence to Control: Probing and Modulating Self-Reflection in Language Models

Fixing What Fine-Tuning Breaks: A Simple and Efficient Method to Improve Safety Post Domain Adaptation

Improving Visual Discriminability of CLIP for Training-Free Open-Vocabulary Semantic Segmentation

Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models

Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data

Understanding Representation Dynamics of Diffusion Models via Low-Dimensional Modeling

Understanding Task Vectors in In-Context Learning: Emergence, Functionality, and Limitations

Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity

Spatial Discriminability of CLIP for Training-Free Open-Vocabulary Semantic Segmentation

Model Collapse Analysis and Improvement for Rectified Flow Models

Understanding Diffusion-based Representation Learning via Low-Dimensional Modeling

GENERALIZATION, ROBUSTNESS AND ADAPTABILITY OF PROGRESSIVE NEURAL COLLAPSE