About
I study how multimodal models use evidence to understand intent and emotion. My interests are in multimodal reasoning, intent recognition, and affective computing.
I am a final-year Electrical and Computer Engineering student at CUHK-Shenzhen and an undergraduate researcher in the NAIS Lab. My lab PI is Prof. Fangxin Wang, and my research mentor is Dr. Dayou Zhang.
Through my work on MISID and AffectiveArt, I have worked on annotation, model evaluation, and multimodal systems. I am interested in understanding where models fail and how to evaluate their reasoning more carefully.
I expect to graduate in June 2027 and am seeking PhD opportunities for Fall 2027.
Co-first author · Under review at AAAI 2027
MISID: Multimodal Intent Recognition
NAIS Lab, CUHK-Shenzhen · Sep 2025 – Present
A dataset and framework for studying hidden intent in long, multimodal conversations in strategic deception games.
- Co-developed the annotation scheme for 3,962 aligned audiovisual utterance segments and independently evaluated inter-annotator agreement.
- Designed FRACTAM's modal decoupling and text reconstruction component for annotation and inference.
- Contributed to model evaluation, conducted ablation studies, and co-designed the LLM-as-a-Judge protocol and FRACTAM framework.
The full framework improved mean Hidden Intent Inference Score from 34.22 to 42.79 across ten language-model backbones.
Method and evaluation details
The modal decoupling component pairs audio with eGeMAPS features and sampled video frames with FACS descriptors. It generates modality-specific descriptions without additional speech transcripts to limit textual interference and support subsequent cross-modal inconsistency analysis.
Inter-annotator agreement reached Krippendorff's α of 0.80–0.94 across structured fields. I contributed to the zero-shot evaluation of 19 language and video-language models. FRACTAM combines modality decoupling, long-range retrieval, and cross-modal evidence chains.
Team lead · 5th of 14 teams
AffectiveArt: Artwork Emotion Understanding
ACM Multimedia 2026 Grand Challenge, Track 2 · May – Jun 2026
A system for recognizing emotion in artworks and generating emotion descriptions.
- Led a three-member team and designed the overall system architecture.
- Fine-tuned image classifiers and designed validation-weighted voting, which outperformed the best individual model.
- Independently fine-tuned SigLIP2 and DINOv3 for visual-attribute salience prediction; salience and classification outputs conditioned a LoRA-fine-tuned Qwen3.5-27B description model.
Models and results
The classifiers used Gemma, Qwen, SigLIP2, and DINOv3. The system achieved 94.0% arousal, 88.3% valence, and 79.2% emotion accuracy on the hidden test set. Generated descriptions are outputs, not inputs to the classifiers.
Course project · Project lead
Interpretable Butterfly Classification
ECE4513 Image Processing and Computer Vision · Summer 2025
Led a four-layer CNN project to classify ten visually similar butterfly species, achieving 83.23% test accuracy. Implemented Class Activation Mapping from scratch and used PCA of feature maps to examine the image regions supporting predictions.
View code on GitHub ↗
Education
The Chinese University of Hong Kong, Shenzhen
B.Eng. in Electrical and Computer Engineering
Computer Engineering stream
Sep 2023 – Jun 2027 (expected)
- Cumulative GPA
- 3.81 / 4.00
- Major GPA
- 3.90 / 4.00
- Major GPA rank · CE stream
- 3 / 60 · Top 5%
Selected coursework (all A): Machine Learning; Machine Intelligence and Applications; Image Processing and Computer Vision; Parallel Programming.
- Dean's List, School of Science and Engineering: 2023–24, 2024–25, 2025–26.
- Academic Specialty Entrance Scholarship, Second Class: 2023 (renewable).
Teaching
Undergraduate Student Teaching Fellow
MAT1001 Calculus I, CUHK-Shenzhen · Sep – Dec 2024
Supported weekly tutorials and office hours for approximately 60 students.
Skills & languages
- Programming
- Python, C++
- ML & vision
- PyTorch, OpenCV
- Tools
- Git, Linux, LaTeX
- Languages
- Mandarin (native); English (IELTS Academic 7.0)