Building a Precise Video Language with Human-AI Oversight (CVPR’26 Highlight, Top 3%)

Zhiqiu Lin, Chancharik Mitra, Siyuan Cen, Isaac Li, Yuhan Huang, Yu Tong Tiffany Ling, Hewei Wang, Irene Pi, Shihang Zhu, Ryan Rao, George Liu, Jiaxi Li, Ruojin Li, Yili Han, Yilun Du, Deva Ramanan

• We propose a cinematic video specification based on 200+ visual primitives to bridge the gap between AI generation and professional content creation.

• We show that the Critique-based Human-AI (CHAI) oversight framework allows an 8B model to outperform GPT-5 and Gemini-3.1-Pro in video reasoning.

• We demonstrate that training on CHAI triplets enables professional-grade generation of complex cinematic techniques like dolly zooms and rack focus.

Learning More by Seeing Less: Structure First Learning for Efficient, Transferable, and Human-Aligned Vision

Tianqin Li*, George Liu*, Tai Sing Lee

• We propose structure-first learning, a pretraining paradigm that uses line drawings to induce more compact and generalizable visual representations.

• We show that training on line drawings develops stronger shape bias, significantly improving data efficiency for tasks.

• We demonstrate that structural pretraining produces compressible representations that lead to better distillation and higher performance in lightweight student models.

3-D Image Based Deep Learning for Dementia Diagnosis

George Liu, Guillermo Goldsztein

• We propose a 3-D convolutional architecture that processes full MRI volumes to capture complex structural patterns missed by traditional 2-D analysis.

• We show that training on T1-weighted MRI scans from the OASIS-3 dataset enables the effective identification of brain atrophy for automated diagnosis.

• We demonstrate that maintaining spatial dimensionality through 3-D convolution significantly improves diagnostic reliability compared to slice-based classification methods.