Haoyu Hu

We introduce a controlled experimental framework for collaborative vibe coding and compare human-led, AI-led, and hybrid groups across 20 experiments with 737 participants. The results show that people contribute uniquely effective high-level guidance, while AI-provided instructions can drive performance collapse. Hybrid systems perform best when humans lead instruction and AI supports evaluation, highlighting the continued importance of human guidance in future hybrid societies.

Across two large-scale dermatology experiments with lay users and primary care physicians, AI assistance improved accuracy and reduced disparities across skin tones. At the same time, multimodal LLM explanations showed a double-edged effect: they increased automation bias for the general public, while experienced clinicians remained more resilient and still benefited even when AI was wrong. The study clarifies how expertise and decision timing reshape human-AI collaboration in medical diagnosis.

MindLLM combines a subject-agnostic fMRI encoder with an off-the-shelf large language model to decode brain activity into text across multiple subjects and tasks. The model improves downstream task performance, generalization to unseen subjects, and adaptation to new tasks, while its neuroscience-informed attention mechanism offers interpretable insight into how semantic information is encoded from fMRI signals.

D-CoRP is a differentiable plugin for refining noisy functional brain network connectivity with an information bottleneck objective. It denoises edges while remaining flexible enough to plug into existing graph neural network backbones, leading to stronger performance across multiple datasets and demonstrating a practical path for more reliable brain-network analysis.