Haojie Huang

I'm an undergraduate student majoring in Artificial Intelligence at Zhejiang University (entered in 2025).

Since February 2026, I have been a research intern at the Learning and Vision Lab, National University of Singapore (NUS@LV-Lab), advised by Prof. Shuicheng Yan and Dr. Xiaobin Hu. At LV-Lab, I have also been very fortunate to work with Xinlei Yu and Chengming Xu.

Email  /  Scholar  /  Github

Research

I'm broadly interested in how vision should be learned, represented, and scaled. I tend to see most problems through representation learning, shaped by three things: the objective, the data, and the architecture. Questions I'm currently working on include: what visual information survives different training paradigms (contrastive pretraining, VLM fine-tuning, native multimodal training); whether pixel-level and semantic objectives can coexist in one representation; and, more recently, how to make vision a first-class citizen in VLM training, one that coexists well with language.

News

• Evidence-RL accepted to NeurIPS 2026.

Publications

Counterfactual Evidence Disentanglement (CED) rewards VLMs during RL post-training only when their answers causally depend on the right visual evidence, improving grounded reasoning across nine benchmarks with no inference-time cost.

A survey of latent space as a computational substrate for language-based models, organized by foundation, evolution, mechanism, ability, and outlook.

* Equal contribution   † Corresponding author