Undergraduate researcher at the University of Texas at Dallas since August 2022, with Phi Kappa Phi and Collegium V Honors. His research centers on language models — how they represent meaning, where they generalize, and where they break — with broader interests connecting machine learning to questions about cognition and scientific discovery. He has held a DAAD RISE research internship in Germany and a policy-oriented internship at CSET, alongside applied work in computer vision, software engineering, and ML for materials and scientific text. He teaches and grades introductory CS courses and is active in student organizations and outreach.

Research Interests

  • Language models and the structure of meaning they learn — including representation analysis, generalization, and behavioral evaluation of frontier systems.
  • Connections between large language models and human language processing, drawing on tools from computational linguistics and cognitive science.
  • Adjacent interests include machine learning for scientific and patent text, ML for materials science, and fairness and representation evaluation in high-stakes AI systems.

Honors and Awards

  • 2026 DAAD RISE Scholarship — $5,000
  • 2026 Undergraduate Research Scholarship Award (URSA) — $500
  • 2026 Spring Archer Fellow — $10,000
  • 2025 Barry Goldwater Scholarship — $7,500
  • 2025 Undergraduate Research Scholarship Award (URSA) — $500
  • 2024 Undergraduate Research Apprenticeship Fellowship (URAP) — $4,500
  • 2024 People’s Choice Award, Association for Computing Machinery Projects
  • 2024 Leland Dillaha MSEE ’97 Undergraduate Scholarship — $1,000
  • 2022 National Merit Scholarship, UT Dallas — $3,000

Languages

  • Japanese — JLPT N1 (150/180; CEFR C1), self-taught.

Places

University of Texas at Dallas

Richardson, TX — August 2022–Present; expected graduation: Dec 2026

B.S. in Computer Science with Honours

Minors in Mathematics and Political Science. GPA overall 3.966/4.0; CS-only GPA 3.96/4.0. Honors: The Honor Society of Phi Kappa Phi, Collegium V Honors, and Dean’s List (x3).

Selected coursework at UT Dallas (titles as on transcript)

  • CS 4V98 — Undergraduate Research in Computer Science
  • CS 4485 — Computer Science Project
  • PSCI 4V76 — Archer Center Washington Internship
  • CS 4349 — Advanced Algorithm Design and Analysis
  • CS 4384 — Automata Theory
  • CS 4337 — Programming Language Paradigms
  • CS 4347 — Database Systems
  • CS 3377 — Systems Programming in UNIX and Other Environments
  • CS 4341 — Digital Logic and Computer Design
  • CS 4141 — Digital Systems Laboratory

Waseda University

Study abroad: academic year 2024–2025 · Tokyo, Japan

All coursework was taught in Japanese (readings, lectures, and assessments).

Graduate-level courses

  • Computer vision
  • Natural language processing
  • Advanced wireless access
  • Distributed, embedded, and real-time processing

Other courses

  • Artificial intelligence and logic
  • Artificial intelligence and complex systems
  • Operating systems
  • Software engineering
  • Data mining

Refereed abstracts

2026

J. V. Le
Companion Proceedings of SPLASH '26, ACM Student Research Competition, pp. 45–46. Two-page extended abstract; peer-reviewed. DOI ACM Showcase on Kudos Play the visual novel

Workshop papers

2026

Attributing Gains in Paired Training: Objective or Data Direction?
J. V. Le, Tien Nguyen
Third NeurIPS Workshop on Attributing Model Behavior at Scale: Data Attribution and Provenance (ATTRIB 2026). 43 of 83 submissions accepted (51.8%). Non-archival.

Preprints

2026

Directional Collapse in Fine-Tuning: Preserving the Way Back
J. V. Le, Tien Nguyen
Submitted to ACL Rolling Review (ARR), October 2026.
Correct Values Can Be Recovered Before Wrong Reasoning Steps
J. V. Le, Tien Nguyen
Submitted to ACL Rolling Review (ARR), October 2026.
Composing Obfuscation Specialists by Merging
J. V. Le, Tien Nguyen
Submitted to ACM International Conference on the Foundations of Software Engineering (FSE).
Router-Free Modular Storage for Knowledge Unlearning in Large Language Models
Vincent Hanke, J. V. Le, Adam Dziedzic, Franziska Boenisch
Submitted to ICLR.
Fast Errors or Slow Effort? Divergent Dual-Process Signatures of Code Obfuscation in Humans and Language Models
J. V. Le, Ilse Lahnstein Coronado, Anh Nguyen, Tien Nguyen
Submitted to ACM International Conference on the Foundations of Software Engineering (FSE).
Extreme Summarization of Worldwide Scientific and Patent Document Clusters for Landscape Analysis
J. V. Le, Rebecca Gelles, Catherine Aiken, Katherine Quinn
Submitted to Scientometrics.
J. V. Le†, Anh Nguyen†, Tien Nguyen
arXiv:2606.31725 [cs.SE]. Submitted to ACM International Conference on the Foundations of Software Engineering (FSE). † Co-first authors (equal contribution). Play the visual novel
A. H. N. Nguyen†, J. V. Le†, Ilse Lahnstein Coronado, T. N. Nguyen
arXiv:2603.07668 [cs.SE]. † Co-first authors (equal contribution). Play the visual novel

Philosophical preprints

2024

Code as Embedding as AI
K. Tanaka Ishii, J. V. Le
Submitted to Minds and Machines.

Press and mentions

2026

ACM Showcase on Kudos · Oct 2026
My SPLASH '26 Student Research Competition abstract, selected by ACM for its research showcase on Kudos.
Emerging Technology Observatory (Substack) · May 19, 2026
Policy post on high-growth AI and cybersecurity clusters in the Map of Science I contributed to.
Emerging Technology Observatory (Substack) · Apr 27, 2026
Map of Science product update (ETO / CSET); related to my work with the data team.
Department of Materials Science and Engineering, UT Dallas · 2026
Recognizes my 2025 Goldwater Scholarship and my summer 2024 research in Dr. Julia Hsu’s group.

2025

Erik Jonsson School of Engineering & Computer Science, UT Dallas
Barry Goldwater Scholarship profile (summer 2025 archive).
UT Dallas News Center · May 28, 2025
Central campus news coverage of 2025 Goldwater Scholars.

2022

Garland ISD Newsroom · May 27, 2022
Named a 2022 National Merit Finalist (Garland High School).

2026

Towards Verifiable LLM Outputs for Software Development
Gulf Coast Undergraduate Research Symposium, Rice University, Oct 2026.
Exact Unlearning for Large Language Models
DAAD RISE Research Symposium.

2025

Head Shape Analysis Using Custom Feature Extraction
UT Dallas Undergraduate Research Scholars Awards Symposium.

2024

Evaluating Large Language Models in Scientific Research: Practical Applications and Challenges
UT Dallas Summer Platform for Undergraduate Students.

2023

Coding Semantics: Sentiment Analysis for Movie Reviews
[Co-presented.] Association for Computing Machinery UT Dallas Research Conference.

2026

UTD CS 2340 — Computer Architecture
Grader
Fall 2026

2025

UTD CS 1337 — Computer Science I
Grader
Fall 2025

2024

UTD CS 2336 — Computer Science II
Grader
Spring 2024
UTD CS 1337 — Computer Science I
Teaching Assistant
Spring 2024
ICODE School — Introduction to Web Development
Technical Lead
Fall 2023 – Spring 2024
Jeremy E. Watson Educational Center
Japanese Instructor
Fall 2023 – Spring 2024

2023

UTD CS 1334 — Programming Fundamentals for Non-Majors
Teaching Assistant
Fall 2023
Garland ISD — Various mathematics courses
Substitute teacher
Fall 2023 – Present

Research

  • Undergraduate Researcher, Dr. Tien Nguyen — Attention reallocation for LLMs in software engineering; Fall 2025 – present.
  • Intern Researcher (DAAD RISE), SprintML, CISPA Helmholtz Center — Efficient exact unlearning for LLMs; Summer 2026.
  • Research Assistant, Dr. Kumiko Tanaka-Ishii — Recurrence quantification analysis of frontier AI models; Fall 2024 – Summer 2025.
  • Undergraduate Researcher, Dr. Ovidiu Daescu — Computer vision and feature selection (cranial shape analysis); Spring 2024 – present.
  • Research Intern, Hsu Lab — ML for materials science; Summer 2024.
  • Student Researcher, Dr. Doug DeGroot — Sentiment analysis for movie reviews; Fall 2023.

Industry

  • CSET, Georgetown University — Intern, Spring 2026.
  • Onplanetz — AI Engineer Intern; Fall 2024 – Spring 2025.
  • Relativity, AI Studio Fellow — Bias and representation in frontier models; Fall 2024.

Other experience

  • UT Dallas Office of Research — Research Information Systems; Summer 2024.
  • Break Through Tech AI — AI/ML Fellow, May 2024 – May 2025.
  • ACM UTD — Back-end developer, Feb 2024 – Apr 2024.
  • UTDesign — Frontend software developer, Jan 2024 – May 2024.
  • Wolters Kluwer Tax and Accounting — Extern, Mar 2023.

Community

  • UT Design — Code review volunteer, Summer 2024.
  • UTD Society of Asian Scientists and Engineers — Secretary and Chair, Jan 2023 – Aug 2024.
  • UTD Artificial Intelligence Society — Officer, Oct 2022 – May 2023.

My library

Ways to see the library 18 views, each opens in a new tab · scroll →
Reading Room

The library as an isometric room: twelve real bookcases, every spine in shelf order, trophies and game boxes where they really sit. Walk up and press E.

Terminus

Mini Metro built from one real library: thirteen bookcases are the stations and every passenger is a book trying to get home to its shelf.

Spine Wall

The whole library as one long wall of shelves, every spine drawn to size in its real place. Zoom in to read the titles; search shines a torch on the book.

Sunburst

The whole library as one ring: genres inside, 889 books on the rim. Click a wedge to dive in, or fold the same tree into a treemap or an icicle.

Strata

The library cut like a cliff face: every book a grain of sediment in the year it was first published, from Gilgamesh to this year.

Constellations

The library's 529 authors as a night sky: stars sized by their books, joined by the shelves, genres, publishers and decades they share.

Flows

The library as 889 threads braided between facets. Pick language, type, genre, bookcase, era or author as columns and watch the ribbons reshape.

By the Numbers

One library weighed, measured and timed: 889 books as 360 kg of paper, a six-storey stack, fifteen years of reading, every count drawn live from the catalogue.

Runs

Every numbered series in the library as a shelf of spines: filled ones are here, dashed ones are the gaps, and a shopping list names what is missing.

Card Catalog

The library as an oak card catalogue: pull a drawer and riffle through 889 typed 3×5 cards filed by author, title, subject, shelf or year.

Stacks

The library as a Wolfenstein-style raycaster: thirteen aisles of real spines drawn from the catalogue. Walk the stacks and pull a book.

Library Orrery

The library as a solar system: every book a planet, the older it is the farther out and the slower it crawls. Scrub ten thousand years for conjunctions.

Shelf Genome

The library as one chromosome: genre bands, year glow, collision hotspots, and fake crossover books.

Read-Order Oracle

Ride the library: twelve genre lines, interchanges where books overlap, and a journey planner for what to read next.

Library Roguelike

A daily text roguelike where every room is a real shelf of the library and every monster is a book.

Bookshelf poster

A whole book collection as one printable SVG poster of spines, sized to the paper you pick.

Bookshelf Tetris

Falling pieces are books. A shelf only clears when its titles read in alphabetical order.

Virtual library

A library catalog as shelves of spines: find a book, browse stats, run integrity checks.

Small things built for fun. Each one is a plain web page with no accounts, no servers and no tracking beyond the site's own. They open in a new tab. Source for the older ones is in NietZteiN/side-projects; the newest live in this site's own repository.

Stories and the site 6

Kamishibai Printer

Prints a Paper Theatre story as kamishibai boards: the picture on each front, and on each back the lines for the board in front of it, so it can be read from behind the stack.

Theatre Studio

A writing room for Paper Theatre scripts: the .vn text on the left, the live stage on the right, lint and the branch walk on every keystroke. Drafts stay in this browser.

Paper Theatre

Every paper and post as a two-minute visual novel: a bet, a result, a citation. One engine; each story is a plain-text script you can read in the repo.

Phonograph

Every blog post pressed onto a twelve-inch record: the groove is the text, paragraphs are tracks, and a tonearm reads it aloud. Drop the needle anywhere.

Blog Post Speedrun

Speedrun the blog: a LiveSplit-style timer, splits, fake unverified world records, a local leaderboard and TAS detection.

Latentland Map

An inked map of Latentland: read the neuron's memoir place by place, or explore it on a budget of surprise.

Machine learning, live 3

Tokenizer Aquarium

Live BPE tokenisation as a fish tank: common tokens swim big and slow, rare ones small and twitchy.

Grokking microscope

A one-layer transformer learning addition mod 113. Scrub through training and watch it grok.

Fake paper titles

A Markov chain that invents ML paper titles, and a linter that scores how cliché a real one is.

Code and obfuscation 5

Site Reverse Engineer

The site's own JavaScript, obfuscated live with the Playground's transforms, plus a daily guessing game.

Ghost Commit Player

Any public GitHub repo replayed as a silent film: intertitles for commits, the file tree growing between.

Obfuscation playground

Stack JavaScript obfuscation transforms into a ladder and diff each rung against the last.

Obfuscation game

Predict the output of obfuscated Python snippets. Tracks correctness, time and confidence.

Git film timeline

A git history laid out like a video editor, with a trailer that plays the whole thing in 30 seconds.

Shows a bundled example repository.

Language 9

Prose Weather

The 10 o'clock forecast for a piece of writing: sentiment is temperature, commas are drizzle, a 45-word sentence is a run-on warning, and every number opens the sentence that caused it.

Grammar Fossil Dig

Paste English and dig down through its word strata: modern loans, Latin, Norman French, Norse, Old English bedrock.

Homophone Bridge

Cross a river of six languages on planks that sound alike; score for how far the meanings drift.

Pidgin Generator

Collide two languages into a toy pidgin: lexicon, phonology, grammar sketch, phrasebook and corpus, all live.

Zipf Karaoke

Words fade by how common they are, then get sung: common ones vanish in a beat, rare ones ring and linger.

Tide Table for Prose

A clock whose face is a paragraph: one word turns into a synonym every minute until midnight.

Poetry form checker

Paste a poem and get morae, syllables, meter and rhyme annotated line by line.

Translation aligner

A source text and its translation side by side, sentence by sentence, with manual realignment.

Aozora frequency reader

Sakaguchi Ango's 堕落論 with the words he uses unusually often highlighted against general Japanese.

Library 12

Library Orrery

The library as a solar system: every book a planet, the older it is the farther out and the slower it crawls. Scrub ten thousand years for conjunctions.

Marginalia Séance

A talking board whose spirits are the authors on these shelves. Ask; the planchette spells an answer stitched from their own book blurbs.

Stacks

The library as a Wolfenstein-style raycaster: thirteen aisles of real spines drawn from the catalogue. Walk the stacks and pull a book.

Bookshelf Planetarium

The real Texas sky right now, each of the 88 constellations renamed for a day after an author on the shelf. Click one for its myth.

Shelf Genome

The library as one chromosome: genre bands, year glow, collision hotspots, and fake crossover books.

Read-Order Oracle

Ride the library: twelve genre lines, interchanges where books overlap, and a journey planner for what to read next.

Book Exchange

Every library book is a listed stock; prices move when its words show up on the blog. Paper-trade 10,000 credits.

Library Roguelike

A daily text roguelike where every room is a real shelf of the library and every monster is a book.

Bookshelf Tetris

Falling pieces are books. A shelf only clears when its titles read in alphabetical order.

Virtual library

A library catalog as shelves of spines: find a book, browse stats, run integrity checks.

Runs on an invented sample catalog.

Bookshelf poster

A whole book collection as one printable SVG poster of spines, sized to the paper you pick.

Harvard Classics oracle

Ask a question or name a mood and get a passage from the Five-Foot Shelf. Also today's Fifteen Minutes a Day reading.

Searches a sample of 340 passages from seven works.

Art and sound 7

True Shuffle

A music app over your YouTube playlists: browse by artist, genre, mood and the anime or game a song is from, then shuffle it properly. A demo runs without signing in.

Signing in works for the owner's Google account; anyone else brings a Google client id of their own.

New
Haruiro

A Japanese poem by a watercolor shoreline: switch between the art print and a reading view, or download the A3 print.

Handwriting Foundry

A tiny type foundry. Draw each letter in its cell, watch the page rewrite itself in your hand, and download a real .otf you can install.

Babel Hexagon

Borges's Library, browsable and complete: every 40-line page that could ever be written sits on a shelf, and whatever you type, the search computes where it already is.

Spectrogram Painter

Paint on a spectrogram and hear it: pitch runs up, time runs right, brightness is loudness. Stamp a word and hide it in the sound like Aphex Twin did.

Text Tartan

Weave any text into cloth: every word a stripe, the first sentence the sett, the whole text the weft of a 2/2 twill you can zoom into thread by thread.

Palimpsest

A vellum leaf that remembers your hand: where you rub, an older hidden text shows through, across visits.

Clocks and maps 4

Passport Atlas

A globe of the eight countries I've set foot in. Spin it, click a lit country, and an ink stamp thumps into the passport beside it.

Dead Authors' Local Time

A world clock for the places where the authors on the shelves died, sorted by whose sun is setting right now.

Haiku clock

The current time, told as a haiku in four languages.

Daylight clock

Local time, sunrise, sunset and daylight arcs for three cities at once.