Long Vu

I’m a 3rd-year Ph.D student in Computer Science and a machine learning researcher who enjoys both the theory and the engineering side of AI.

On the theory side, I’m interested in how models reason, why they can extract patterns, make predictions, and generalize across tasks. My current research on this side explores in-context learning for time-series forecasting, using ideas from Bayesian inference and Prior-Data Fitted Networks.

On the practical side, I love tinkering with systems and hardware. I like working close to the profiling, probing, debugging, and tuning models for on-device deployment. Lately, I’ve been focusing on making large models run efficiently on-device, especially as specialized NPUs and edge hardware become more common.

Foundation Time-Series Model with In-Context Learning

Designing compact models capable of Bayesian in-context inference

• Designing a compact time-series foundation model capable of state-of-the-art zero-shot forecasting, trained on small synthetic datasets with controlled temporal structure.
• Improved the model’s ability to generalize across real-world series by designing efficient temporal representations, tailoring synthetic priors, and optimizing training stability and inference efficiency.

Efficient Large Model Architectures for On-Device Deployment

Understand and design compact, hardware-aware models for real-world edge environments

• Investigating how model architecture and design influence efficiency, robustness, and representational capacity when deployed on constrained hardware (CPU/GPU/NPU).
• Applying neural architecture search and structural analysis to design compact, production-ready models, with attention to energy consumption, memory footprint, and latency under real deployment conditions.
• Studying decision boundaries and failure modes in small-scale models to understand what architectural features support stable and reliable inference in on-device LLMs and vision models.