Hieu Pham Dinh

Hi! I’m Hieu, a Data Science undergraduate at VinUniversity.

My research focuses on bridging the gap between robotics research and real-world deployment. I am interested in developing robotic systems that can understand their environments, reason about complex tasks, and act reliably outside controlled laboratory settings.

To this end, I work on multimodal robot learning, vision-language-action models, agentic AI, and reinforcement learning for decision-making and control. I am particularly interested in building systems that integrate perception, reasoning, and action to enable more capable and adaptable robots.

Previously, I have worked on projects spanning robotics, medical AI, intelligent control, and natural language processing, with experience in dataset development, model training, evaluation, and deployment-oriented research.

I am always happy to connect for research discussions and collaborations.

news

Aug 18, 2026	2 papers accepted at MICCAI 2026 workshops: UNSURE Workshop and FAIMI/BRIDGE/EPIMI Workshop!
Feb 01, 2026	Started as a Research Intern at Knovel Engineering, working on multimodal learning and vision–language–action modeling in medical and surgical contexts.
Dec 01, 2025	Presented a poster at Open Innovation Conference 2025, Workshop on Digital Twin for Greener and Healthier Indoor Spaces.
Nov 01, 2025	Received a Consolation Prize in the Regional Round of the AI Olympiad for Undergraduate Students, including a top-two national performance in the Word Translation task.
Sep 01, 2025	Named to the VinUniversity Dean’s List for Academic Excellence for AY 2024–2025.
Aug 01, 2025	Became Co-President of VinTelligence, VinUniversity’s AI and Data Science club.

latest posts

Jun 21, 2026	AI Looks Easy Until You Put It in a Robot
Jun 21, 2026	Why Everyone Is Talking About World Models — Thoughts From Someone Without 100,000 GPUs

selected publications

UNSURE ’26

Beyond Uncertainty: Generalizable Failure Monitoring for Surgical Segmentation under Acquisition Degradation

Hieu D. Pham, Dang P. M. Cao, and Thanh Trung Huynh

In UNSURE Workshop, MICCAI, 2026

Abs arXiv Code

We propose TCSR-Monitor, a post-hoc failure monitoring framework that goes beyond uncertainty estimates by incorporating complementary observable signals such as shape, temporal consistency, and image quality. The framework operates without access to model internals or ground truth during deployment and generalizes to unseen acquisition degradations, helping make AI systems more reliable in real surgical environments.
FAIMI/BRIDGE/EPIMI ’26

When Oracle Conditioning Misleads Deployment: Conditioning-Availability Bias in Echocardiographic Segmentation

Dang P. M. Cao, Hieu D. Pham, and Hieu Pham

In FAIMI/BRIDGE/EPIMI Workshop, MICCAI, 2026

Abs arXiv Code

This work highlights a deployment issue that often goes unnoticed: models are frequently evaluated using auxiliary information that will not actually be available in practice. We introduce a simple audit framework to expose this conditioning-availability bias, showing how oracle-conditioned evaluation can substantially overestimate real-world performance and fairness.
LREC ’26

VietJobs: A Vietnamese Job Advertisement Dataset

Hieu Pham Dinh, Hung Huy Nguyen, and Mo El-Haj

In Language Resources and Evaluation Conference, 2026

Abs arXiv Code

VietJobs is a large-scale corpus of 48,092 Vietnamese job advertisements containing more than 15 million words and spanning all 34 provinces and municipalities of Vietnam. It includes structured information such as job titles, categories, salaries, skills, employment conditions, occupational domains, and work types, supporting research in Vietnamese NLP and labour-market analytics.
LREC ’26

Structured Prompting for Arabic Essay Proficiency: A Trait-Centric Evaluation Approach

Salim Al Mandhari, Hieu Pham Dinh, Mo El-Haj, and Paul Rayson

In Language Resources and Evaluation Conference, 2026

Abs arXiv Code

This work introduces standard, hybrid, and rubric-guided prompting strategies for trait-specific Arabic automatic essay scoring. Eight language models are evaluated on the QAES dataset across organization, vocabulary, development, and style in zero-shot and few-shot settings.