Portrait of Tung Duong Ta
Graduating Dec 2026 · Open to research & ML roles from early 2027

Tung Duong Ta

Ph.D. Candidate in Computer Science · CoRaL Lab, University of Maryland, Baltimore County

I build AI systems that read what is hard to read: handwritten Vietnamese, 3D CT scans, and multi-step math problems.

CV (PDF) Google Scholar GitHub LinkedIn dta1@umbc.edu

About

I am a Ph.D. candidate in Computer Science at the University of Maryland, Baltimore County (UMBC), advised by Dr. Tim Oates in the CoRaL Lab. I was admitted to candidacy in February 2026 and will graduate on December 18, 2026 (GPA 3.933/4.0).

My dissertation studies low-resource Vietnamese handwriting recognition: a new three-level dataset and a multi-path architecture with fusion that generalizes to 3D CT report generation and mathematical reasoning of large language models.

Before my Ph.D., I was a Big Data Engineer at CMC Corporation in Hanoi, where I built knowledge graph search and fine-tuned BERT models for Vietnamese NER and text-to-SQL. I hold a B.S. in Computer Science (Magna Cum Laude) from Villanova University.

I am open to full-time AI Research Scientist, Applied Scientist, and Machine Learning Engineer roles starting in early 2027, in the United States or Vietnam.

Research interests

Vietnamese handwriting recognition Vision-language models Stroke modality Deep ensembles 3D CT report generation LLM mathematical reasoning ICU outcome prediction Raman spectroscopy Knowledge graphs

Dissertation

“Low-Resource Vietnamese Handwriting Recognition: A Three-Level Dataset and a Multi-Path Architecture with Fusion Generalized to Computed Tomography Report Generation and Mathematical Reasoning”

Advisor: Dr. Tim Oates · Committee: Dr. Lara Martin, Dr. Justin Brooks, Dr. Manas Gaur, Dr. Trung Vu

News

  • Ph.D. defense and graduation from UMBC (expected December 18, 2026).
  • MDEF accepted as a regular paper at the LLM4MDF Workshop at IEEE ICDM 2026 (CORE A*). Code
  • Co-authored paper on hierarchical Raman classification of skin lesions submitted to Skin Research and Technology (Wiley); now under review.
  • PIKE submitted to AAAI 2027.
  • Started as a Grader for Software Engineering I at UMBC.
  • MDToC published in the Journal on Information Technologies & Communications (ICT Research), Vol. 2026, No. 2.
  • Earned the M.S. in Computer Science from UMBC en route to the Ph.D.
  • Admitted to Ph.D. candidacy.
  • MDToC preprint released on arXiv:2512.18841.

Publications

Full list on Google Scholar

Conference & workshop papers

[C1]
MDEF: Multi-Report Deep Ensemble Fusion for 3D CT Report Generation
Tung Duong Ta, Tim Oates, Manas Gaur, Trong-Nghia Nguyen, Tien-Cuong Nguyen, Tuan-Cuong Vuong, Trang Xuan Mai, Thien Van Luong
AcceptedWorkshop on Large Language Models for Multimodal Data Fusion (LLM4MDF) at the IEEE International Conference on Data Mining (ICDM 2026, CORE A*). Regular paper.
Code

Journal articles

[J1]
MDToC: Metacognitive Dynamic Tree of Concepts for Boosting Mathematical Problem-Solving of Large Language Models
Tung Duong Ta, Tim Oates, Thien Van Luong, Huan Vu, Tien Cuong Nguyen
PublishedJournal on Information Technologies & Communications (ICT Research), Vol. 28, No. 2 (Vol. 2026, No. 2), pp. 59–68, September 2026.
PDF Abstract Journal issue (pp. 59–68) arXiv Code

Under review

[R1]
PIKE: Patient-Context Informed Knowledge Graph Enrichment for ICU Outcome Prediction
Tuan-Cuong Vuong, Trang Mai Xuan, Son Thai Mai, Tung Duong Ta, Cuong Tien Nguyen, Huan Vu, Vu-Duc Ngo, Thien Van Luong
Under reviewAAAI Conference on Artificial Intelligence (AAAI 2027).
[R2]
Hierarchical Native-Coordinate Raman Classification of Skin Lesions under Provenance-Aware Repeated Validation
Chu Duc Hoang, Nguyen Thanh Tung, Truong Huu Hoang, Tung Duong Ta
Under reviewSkin Research and Technology (Wiley). Submitted October 2026.

In revision

[P1]
ViTLD: A Three-Level Vietnamese Handwriting Dataset from Photographed Pages
Tung Duong Ta, Thai Dang Tran, The-Son Le, Quang Dang, Trung Vu, Manas Gaur, Tim Oates
In revisionIn revision for resubmission.
[P2]
SIME: A Stroke-Image Modality Ensemble in Word-Level Vietnamese Handwriting Recognition
Tung Duong Ta, Thai Dang Tran, The-Son Le, Quang Dang, Trung Vu, Manas Gaur, Tim Oates
In revisionIn revision for resubmission.

Research projects

CoRaL Lab, UMBC · 2023 to present

MDEF: Multi-Report Deep Ensemble Fusion for 3D CT Report Generation

First author · LLM4MDF at ICDM 2026 · Accepted
  • Proposed MDEF, a deep-ensemble pipeline that turns the single-report paradigm into a multi-report paradigm.
  • Outperformed MARCH, the best single-report baseline, by +0.079 and +0.045 CE F1 on RadGenome and CT-RATE, with a small positive margin on INSPECT.
  • Raised recall by +70% to +97% relative to MARCH, averaged over the 18 pathologies.
Code

MDToC: Metacognitive Dynamic Tree of Concepts

First author · ICT Research · Published
  • Proposed MDToC, a three-phase prompting technique built on the metacognitive stages of planning, monitoring, and reviewing.
  • Outperformed GoT by 5% on CHAMP, 5.4% on MATH, and 4% on Game-of-24, three mathematical benchmarks.
  • Outperformed ToT by up to 19% and GoT by up to 13% across the backbone models.
Paper Code arXiv

ViTLD: A Three-Level Vietnamese Handwriting Dataset from Photographed Pages

First author · In revision for resubmission
  • Built ViTLD from 565 paper images of 435 writers, the first Vietnamese handwriting dataset with word, line, and paragraph annotation on photographed papers.
  • Proved ViTLD harder than the two existing three-level datasets, VNOnDB and UIT-HWDB, raising the average word-level CER of nine handwriting recognition systems to 12.56 from 7.32 and 9.39.
  • Proved ViTLD harder at line level, raising the average CER to 22.57, from 14.09 on VNOnDB and 16.91 on UIT-HWDB.

PIKE: Patient-Context Informed Knowledge Graph Enrichment for ICU Outcome Prediction

Contributing author · AAAI 2027 · Under review
  • Contributed to PIKE, built on Evidence-Guided Node Selection (EGNS) and Bidirectional Low-Rank Fusion (BLRF).
  • Outperformed the strongest baseline by 2.2 to 4.0 AUROC points in each setting on MIMIC-III and MIMIC-IV.
  • Improved AUPRC by up to 6.9 points over CTPD, the strongest baseline on MIMIC-IV mortality.

SIME: A Stroke-Image Modality Ensemble in Word-Level Vietnamese Handwriting Recognition

First author · In revision for resubmission
  • Proposed SIME, which adds the stroke modality, reconstructed from the image by stroke derendering, to an ensemble of fine-tuned VLMs.
  • Outperformed the strongest image modality ensemble on two out-of-distribution benchmarks, reducing CER by about 11% on VN-Para and 9% on VN-Addr.
  • Released VN-Para, a diacritic-rich word-level benchmark of 8,042 handwritten word images from 100 documents.

Hierarchical Native-Coordinate Raman Classification of Skin Lesions

Contributing author (data curation, supervision, resources) · Skin Research and Technology · Under review
  • Contributed to H-QFR, a hierarchical classifier (benign vs. malignant, then BCC vs. melanoma) that learns directly on the native, nonuniform Raman wavenumber grid instead of resampling to an artificial uniform axis.
  • Evaluated with a provenance-aware validation protocol using source-family grouped splits (ten repeats of five-fold outer, three-fold inner) to prevent leakage between related spectra.
  • Reached 82.87% repeat-level balanced accuracy and 83.59% overall accuracy on BCC, benign, and melanoma spectra, +1.49 points over the inner-selected baseline.

Skills used across projects

Tools and models
PyTorchHugging Face TransformersPEFT / LoRA Qwen2.5-VLInternVL3TrOCRRadBERT GPT-4-TurboMONAIOpenCVSLURM multi-GPU Distributed training

Earlier projects

Diagram of LLM-generated metacognitive prompts
Auto-metacognitive prompting · UMBC, 2023

Had GPT-3.5-turbo generate metacognitive prompts to solve and evaluate math and logic solutions step by step, improving on the FELM factuality benchmark baseline. This work grew into MDToC.

Hierarchical reinforcement learning diagram
Multi-agent hierarchical RL · UMBC summer research

Surveyed 15 multi-agent and graph RL papers and built a deep RL model for multi-task job scheduling.

Knowledge graph company similarity search
Similar company search · CMC Corporation

Knowledge graph embeddings (TransE, ComplEx) with Random Walk with Restart and HNSW / ScaNN search for a B2B recommendation product.

Natural language to SQL
Vietnamese natural language to SQL · CMC Corporation

Built a Vietnamese transaction dataset and fine-tuned multilingual BERT / RAT-SQL models for real-time text-to-SQL (78% accuracy).

Experience

  1. Graduate Research Assistant

    Jan 2023 – Present
    CoRaL Lab, University of Maryland, Baltimore County · Baltimore, MD · Advisor: Dr. Tim Oates
    • Research on vision-language models, deep ensembles, LLM reasoning, and medical AI.
    • Led MDEF, MDToC, ViTLD, and SIME as first author; contributed to PIKE (see Research).

  2. Teaching Assistant and Grader

    Aug 2022 – Present
    Department of Computer Science and Electrical Engineering, UMBC
    • Teaching Assistant (Fall 2022 – Spring 2026): Natural Language Processing (two terms), Algorithms, Computer Architecture (undergraduate and graduate), Databases, Data Visualization, Mobile Computing.
    • Led lab and discussion sessions, held office hours, and graded assignments and exams.
    • Grader (Fall 2026 – Present): Software Engineering I.

  3. Big Data Engineer

    May 2021 – Jul 2022
    CMC Corporation · Hanoi, Vietnam
    • Implemented Random Walk with Restart on knowledge graph embeddings for company similarity search, with 2% higher accuracy and 5% faster ranking.
    • Fine-tuned BERT-based models for NER-based business information extraction (85% accuracy).
    • Fine-tuned BERT-based models for Vietnamese natural language-to-SQL conversion (78% accuracy).
    • Worked with Neo4j, AmpliGraph, and FAISS to store, embed, and search large company knowledge graphs.

  4. Undergraduate Teaching Assistant

    Jan 2019 – 2021
    Villanova University · Villanova, PA
    • Graded submissions and gave feedback for CSC 1052 and CSC 2053; supported students outside class.

  5. Undergraduate Research Assistant

    Jun 2019 – Aug 2019
    Villanova University · Villanova, PA
    • Wrote a Nest.js back end that tags problems using Stack Overflow's tag structure.
    • Wrote questions and warm-up challenges for the one-up learning website (3% gain in educational effectiveness).

Education

  1. Ph.D. in Computer Science

    Aug 2022 – Dec 2026 (expected)
    University of Maryland, Baltimore County · GPA 3.933/4.0
    • Advisor: Dr. Tim Oates. Committee: Dr. Lara Martin, Dr. Justin Brooks, Dr. Manas Gaur, Dr. Trung Vu.
    • Admitted to candidacy in February 2026. Expected graduation on December 18, 2026.

  2. M.S. in Computer Science

    May 2026
    University of Maryland, Baltimore County · Earned en route to the Ph.D.

    Diploma · Graduate transcript (PDF)

  3. B.S. in Computer Science, Magna Cum Laude

    2017 – May 2021
    Villanova University · GPA 3.81/4.0

    Diploma · Undergraduate transcript (PDF)

  4. High school, Physics specialization

    2014 – 2017
    Hanoi–Amsterdam High School for the Gifted · Hanoi, Vietnam

Honors & awards

Click a certificate to enlarge
  • Magna Cum Laude, B.S. in Computer Science, Villanova University.
  • Bayesian Methods for Machine Learning certificate, HSE University (Coursera).
  • Second Place, CCSC Eastern Region Programming Contest, Consortium for Computing Sciences in Colleges (Villanova “Team Wild” with James Gamboli and Zachary DeStefano).
  • Second Place, ICPC Mid-Atlantic USA Regional Contest, Wilkes University site (26 teams), with Villanova team “Cats01”.
  • Second Place, CCSC Eastern Programming Contest.
  • Consolation Prize in Physics, Hanoi City Excellent Student Competition, Grade 12 (2016–2017), Hanoi–Amsterdam High School for the Gifted.
  • Third Prize in Physics, Hanoi City Excellent Student Competition, Grade 12 level (2015–2016), awarded as a Grade 11 student.
  • Second Prize in Physics, Hanoi City Excellent Student Competition, Grade 9, Doan Thi Diem Secondary School.
CCSC Eastern 20202nd Place, Programming Contest
ICPC Mid-Atlantic 20192nd Place, Wilkes University site
CCSC Eastern 20192nd Place, Programming Contest
Hanoi City Physics 2015–2017Third Prize and Consolation Prize
Hanoi City Physics 2016Certificate of merit, Third Prize
Hanoi City Physics 2014Certificate of merit, Second Prize

Degrees & transcripts

M.S. diploma, UMBCComputer Science, 2026
B.S. diploma, VillanovaMagna Cum Laude, 2021
First page of UMBC graduate transcript
UMBC graduate transcriptGPA 3.933/4.0 · PDF
First page of Villanova undergraduate transcript
Villanova transcriptGPA 3.81/4.0 · PDF

Student ID numbers, date of birth, and home address are redacted from the posted transcripts. Official copies are available on request.

Technical skills

Languages
Python, C++, Java, SQL
Frameworks & libraries
PyTorch, Hugging Face Transformers, PEFT (LoRA), OpenCV, MONAI
ML / AI techniques
Vision-language model fine-tuning, deep ensembles, LLM prompting, knowledge graph embeddings
Models
Qwen2.5-VL, InternVL3, TrOCR, RadBERT, GPT-4-Turbo
Tools & infrastructure
SLURM multi-GPU clusters, distributed PyTorch training, Git, LaTeX, Neo4j, AmpliGraph, FAISS