2026
The Geometric Mechanics of Contrastive Learning: Alignment Potentials, Entropic Dispersion, and Modality Gap
Yichao Cai; Zhen Zhang; Yuhang Liu; Javen Q. Shi.
arXiv preprint 2026
A theoretical study of contrastive learning geometry, alignment forces, dispersion, and the emergence of modality gap.
The Geometric Mechanics of Contrastive Learning: Alignment Potentials, Entropic Dispersion, and Modality Gap
Yichao Cai; Zhen Zhang; Yuhang Liu; Javen Q. Shi.
A theoretical study of contrastive learning geometry, alignment forces, dispersion, and the emergence of modality gap.
2025
I Predict Therefore I Am: Is Next Token Prediction Enough to Learn Human-Interpretable Concepts from Data?
Yuhang Liu; Dong Gong; Yichao Cai; Erdun Gao; Zhen Zhang; Biwei Huang; Mingming Gong; Anton van den Hengel; Javen Q. Shi.
International Conference on Learning Representations (ICLR) 2026
An investigation of whether next-token prediction alone is sufficient for learning human-interpretable concepts from data.
I Predict Therefore I Am: Is Next Token Prediction Enough to Learn Human-Interpretable Concepts from Data?
Yuhang Liu; Dong Gong; Yichao Cai; Erdun Gao; Zhen Zhang; Biwei Huang; Mingming Gong; Anton van den Hengel; Javen Q. Shi.
An investigation of whether next-token prediction alone is sufficient for learning human-interpretable concepts from data.
On the Value of Cross-Modal Misalignment in Multimodal Representation Learning
Yichao Cai*; Yuhang Liu*; Erdun Gao; Tianjiao Jiang; Zhen Zhang; Anton van den Hengel; Javen Q. Shi. (* equal contribution)
Advances in Neural Information Processing Systems (NeurIPS) 2025 Spotlight
Studies when controlled cross-modal misalignment can improve multimodal representation learning instead of only harming it.
On the Value of Cross-Modal Misalignment in Multimodal Representation Learning
Yichao Cai*; Yuhang Liu*; Erdun Gao; Tianjiao Jiang; Zhen Zhang; Anton van den Hengel; Javen Q. Shi. (* equal contribution)
Spotlight
Studies when controlled cross-modal misalignment can improve multimodal representation learning instead of only harming it.
2024
CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts
Yichao Cai; Yuhang Liu; Zhen Zhang; Javen Q. Shi.
European Conference on Computer Vision (ECCV) 2024
Explores language-guided disentanglement of style and content through contrastive learning with augmented prompts.
CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts
Yichao Cai; Yuhang Liu; Zhen Zhang; Javen Q. Shi.
Explores language-guided disentanglement of style and content through contrastive learning with augmented prompts.