Hi, I am Henry Zhu.

OSCAR combines Monte Carlo Tree Search, dual-granularity rewards, and online preference optimization to reduce hallucinations in vision-language models.

Minghui Chen, Chenxu Yang, Hengjie Zhu, Dayan Wu, Qingyi Si, and Zheng Lin. 2026. Online Self-Calibration Against Hallucination in Vision-Language Models. In Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence (IJCAI 2026), 989–997. https://doi.org/10.24963/ijcai.2026/111

KDPH models class proxies with anisotropic Kent distributions to absorb gradient conflicts in multi-label cross-modal hashing.

Hengjie Zhu, Dayan Wu, Zihao Zhang, Xinze Liu, Jingxuan Yu, Peng Fu, Zheng Lin, and Weiping Wang. 2026. Absorbing Gradient Conflicts: Modeling Semantic Variance via Kent Distributions for Cross-Modal Hashing. In Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence (IJCAI 2026), 2145–2153. https://doi.org/10.24963/ijcai.2026/239

This paper rethinks strict discretization for asymmetric image retrieval and proposes Deep Correlation Alignment Hashing (DCAH) with correlation alignment based quantization.

Xinze Liu, Dayan Wu, Hengjie Zhu, Chenming Wu, and Pengwen Dai. 2026. Discretization Is Not Always Better: Rethinking Deep Quantization for Asymmetric Image Retrieval. Proceedings of the AAAI Conference on Artificial Intelligence, 40(9), 7323-7331. https://doi.org/10.1609/aaai.v40i9.37670

This paper proposes a dynamic equilibrium gradient-assigned attack against graph prompt learning, which can successfully attack the graph prompt learning model with a small amount of perturbations.

Ju Jia, Jingxuan Yu, Di Wu, Cong Wu, Hengjie Zhu, and Lina Wang. 2025. Prompt as a Double-Edged Sword: A Dynamic Equilibrium Gradient-Assigned Attack against Graph Prompt Learning. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD 25). Association for Computing Machinery, New York, NY, USA, 1049–1060. https://doi.org/10.1145/3711896.3737091

MulVec combines global and local visual evidence with four semantic retrieval roles for training-free zero-shot composed image retrieval.

Zihao Zhang, Dayan Wu, Xinze Liu, Hengjie Zhu, Yiliang Zhu, Ding Wang, Peng Fu, Zheng Lin, and Weiping Wang. 2026. MulVec: Fine-Grained Role-Aware Matching for Training-Free Zero-Shot Composed Image Retrieval. arXiv preprint arXiv:2608.25305. https://doi.org/10.48550/arXiv.2608.25305

FOVEA retrieves visual evidence on demand and applies a gated correction to draft states for efficient multimodal speculative decoding.

Hengjie Zhu, Dayan Wu, Zihao Zhang, Xinze Liu, Jingxuan Yu, Peng Fu, Zheng Lin, Weiping Wang, and Ding Wang. 2026. FOVEA: Focused On-Demand Visual Evidence Adaptation for Cache-Friendly Multimodal Speculative Decoding. arXiv preprint arXiv:2608.22883. https://doi.org/10.48550/arXiv.2608.22883

Verifier-Guided Decoding uses intrinsic grounding signatures to detect risky object mentions and selectively regenerate them during vision-language generation.

Lei Yang, Xinze Liu, Dayan Wu, Ding Wang, Hengjie Zhu, Zihao Zhang, Tianzhu Hu, Hanqi Wu, Peng Fu, and Zheng Lin. 2026. Hallucinations Leave a Grounding Signature: Verifier-Guided Decoding for Selective Object Correction. arXiv preprint arXiv:2607.27823. https://doi.org/10.48550/arXiv.2607.27823

HashViT learns binary retrieval representations within a Vision Transformer using a dedicated HASH token and a lightweight refinement adapter.

Xinze Liu, Ding Wang, Dayan Wu, Hengjie Zhu, Peng Fu, Zheng Lin, and Weiping Wang. 2026. Beyond Post-Quantization: Native Hash Learning with a Dedicated HASH Token. arXiv preprint arXiv:2607.03328. https://doi.org/10.48550/arXiv.2607.03328