Project Experience
Publication
Where Fact Ends and Fairness Begins: Redefining AI Bias Evaluation through Cognitive Biases
Jen-tse Huang, Yuhang Yan, Linqi Liu, Yixin Wan, Wenxuan Wang, Kai-Wei Chang, Michael R. Lyu
| EMNLP Findings 2025 | DOI: 10.18653/v1/2025.findings-emnlp.583 | code |
Selected Projects
- AI-enabled Digital Health: Wearables & Hepatology Jan. 2026 - present
Supervisor: Prof. Deborah Estrin & Prof. JP Pollak New York, NY, USA- Built a clinician-facing workbench for liver-disease research, combining cohort analysis with longitudinal patient views.
- Developed a data pipeline converting raw wearable-device data into 9 standardized health-data schemas (OMH/IEEE).
- Generated synthetic research data by combining Synthea patient profiles with simulated wearable time series for platform development.
- Developing predictive workflows to discover digital biomarkers and forecast liver-disease progression and patient outcomes.
- Can AI Agent Fit in Human Society? [Slides 1] [Slides 2] Apr. 2024 - Aug. 2025
Supervisor: Prof. Michael R. Lyu & Dr. Jen-tse Huang (ARISE@CUHK) Hong Kong SAR- Built the open-source Fact-or-Fair benchmark to evaluate factuality–fairness trade-offs across 19 indicators and 3 cognitive biases.
- Built evaluation pipelines for 6 LLMs and 4 text-to-image models, analyzing factual responses and gender and racial representation.
- Exposed demographic over-balancing, showing that seemingly fair outputs often misrepresented real group differences.
- Evaluation on the Vulnerability of Current Generative Models [Slides] Feb. 2024 - Jun. 2024
Supervisor: Prof. Sabine Süsstrunk & Dr. Daichi Zhang (IVRL@EPFL) Lausanne, Switzerland- Benchmarked jailbreak defenses across language and text-to-image models, revealing persistent vulnerabilities to adversarial prompts.
- Created and annotated a 2,000-image dataset to quantify gender and skin-tone bias in Stable Diffusion portraits across five professions.
- Evaluated Fair Diffusion for bias mitigation, testing controlled gender and skin-tone representation across professions and model versions.
- Audited fake-image detectors on AI-generated portraits, uncovering substantial performance gaps across demographic groups.
- Efficient Video Analytics [Poster] Jun. 2023 - Sep. 2023
Supervisor: Prof. Eric Lo (CPII@CUHK) Hong Kong SAR- Received the CUHK 2023 Best Project Award, ranking Top 5 among 58 undergraduate research projects.
- Built a multimodal search system for airport lost-and-found using natural-language descriptions or reference images.
- Improved retrieval quality by reducing duplicate detections and ranking the most relevant people and belongings in crowded scenes.
- Delivered an end-to-end prototype with frame and clip retrieval, custom test cases, visualizations, and a React web interface.