LLMPopcorn
An LLM-assisted pipeline for generating and evaluating titles, cover prompts, and short-video prompts designed for audience appeal.
Vision, language, and world models
Multimodal AI
We evaluate and adapt models that connect language, images, video, and product data. The work spans physical reasoning, missing-modality completion, video-generation evaluation, and multimodal recommendation, emphasizing measurable tasks and primary research artifacts rather than demonstration-only claims.
Projects and implementations
4 worksMMPCBench
A benchmark for measuring how multimodal language models reconstruct missing product text or imagery and support recommendation.
PhysicsMind
A simulation-and-real-world benchmark for testing physical reasoning and prediction in vision-language and world models.
IISAN-Versa
A decoupled parameter-efficient adaptation framework for symmetric and asymmetric multimodal foundation models in recommendation.
Publications
5 papersBenchmarking Multimodal Large Language Models for Missing Modality Completion in Product Catalogues
J Fu, W Deng, K Zheng, et al.
EVE-Bench: Entropy-Guided Benchmarking and Evaluation for Frontier Video Generation Models
Nai-Xin Zhai, Qingyuan Hong, Xiao Jin, et al.
LLMPopcorn: Exploring LLMs as Assistants for Popular Micro-Video Generation
J Fu, X Ge, K Zheng, et al.
PhysicsMind: Sim and Real Mechanics Benchmarking for Physical Reasoning and Prediction in Foundational VLMs and World Models
CW Mak, G Zhu, B Zhang, et al.
ContrastiveGaussian: High-Fidelity 3D Generation with Contrastive Learning and Gaussian Splatting
J Liu, E Huang, D Mao, et al.