About Research Projects Honors Contact

Cambridge, Massachusetts

Jia Hao LiangFrank

Artificial Intelligence & Decision Making, Mathematics

I am a sophomore at MIT interested in language models, model architecture, and mathematics.

Email me GitHub Résumé
Portrait of Jia Hao Liang

About

I'm a student at MIT, majoring in Artificial Intelligence and Decision Making (Course 6-4) and Mathematics (Course 18). Before that, I spent a year at Columbia University as a Science Research Fellow, finishing with a 4.16 GPA and Dean's List honors every term, before transferring. My work sits at the intersection of machine learning systems and theory, from fingerprinting multi-agent LLM frameworks against adversarial theft to implementing transformer architectures and distributed training from scratch. Outside research, I compete in math and physics olympiads.

MIT seal

2026 – 2029

Massachusetts Institute of Technology

AI & Decision Making (6-4) & Mathematics (18)

Columbia University coat of arms

2025 – 2026, Transferred

Columbia University, Columbia College

GPA 4.16/4.00 · Science Research Fellow · Dean's List, all terms

Research & Experience

September 2026 – Present

Computer Science and Artificial Intelligence Laboratory (CSAIL), MIT

Student Researcher (UROP), advised by Prof. Vincent Sitzmann and Xingjian Bai

  • Investigating agents as optimizers for neural network training, including evolution strategies, per-layer optimizer code generation, scheduled reviews, and predictive monitoring.

Project page ↗ · GitHub ↗ · On X ↗

2026

Machine Learning Passion Project

Researcher, Massachusetts Institute of Technology

  • Built a Transformer-based language model from scratch: BPE tokenizer, attention blocks, and decoder. Published trained weights on Hugging Face.
  • Profiled forward/backward/optimizer performance, implemented a fused FlashAttention-2 Triton kernel, and distributed training across GPUs with DDP, optimizer sharding, and FSDP.

2025 – 2026

Software Systems Laboratory (SSL)

Researcher, Columbia University, advised by Prof. Junfeng Yang and PhD student Yun-Yun Tsai

  • Investigated fingerprinting methods for distributed agentic frameworks to defend multi-agent LLM systems against adversarial theft.
  • Developed RAFP, a rare-region adversarial fingerprinting method for LLM lineage verification, with theoretical robustness guarantees.

2024

Yau MathCamp

Best Project Award, Group Leader, Fudan University, Shanghai

  • Studied Galois theory and researched the Inverse Galois Problem; selected to present findings at the concluding colloquium (5 of 70).

2023 – 2025

Macau Intelligent Bus Program

Developer, Third Prize, Macau University of Science and Technology

  • Built a mobile app combining real-time bus location with ML and algorithmic optimization to compute efficient commuter routes. Third Prize, Youth AI and Future Media Global Innovation Contest.

Projects

Research Project page ↗ GitHub ↗ On X ↗

OPAL

Reactive Optimizers — OPAL project preview

OPtimizer with an Agent in the Loop

A reactive optimizer that lets a coding agent inspect training dynamics and modify the optimizer as a model trains. The agent can adjust update rules, schedules, and diagnostics within a single run. The accompanying PyTorch package makes this approach available for existing training workflows.

OptimizationLLM agentsPyTorch
Passion Project Open map ↗ Reddit ↗

FindMyCharlie

FindMyCharlie map showing CharlieCard reports across Boston subway stations

A community map for Boston’s collectible CharlieCards

I built FindMyCharlie to help fellow transit enthusiasts find all eight card designs without scrolling through hundreds of Reddit comments. It maps reports by station and machine, combines scheduled Reddit reviews with community submissions, and hides sightings superseded by reports of empty or broken machines.

Page views
—
Community reports
—
Reddit comments reviewed
—

Loading live totals…

Loading latest Reddit review status…

Interactive mapsCommunity reportingPublic transit
Research Preprint on arXiv

RAFP

Rare-region Adversarial Fingerprinting for LLM Lineage Verification

A fingerprinting method that defends a language model against adversarial theft. RAFP embeds identifiable signals in rare regions of the model's input space and comes with theoretical robustness guarantees for verifying model lineage. Developed at Columbia's Software Systems Laboratory under Prof. Junfeng Yang.

Model lineage Model security LLM fingerprinting
Passion Project Weights on Hugging Face

Transformer LM From Scratch

A language model built end to end, then made fast

A Transformer-based language model implemented from the ground up, including a BPE tokenizer, attention blocks, and decoder, followed by systems-level optimization: profiling forward/backward/optimizer passes, writing a fused FlashAttention-2 Triton kernel, and distributing training across GPUs with DDP, optimizer sharding, and FSDP.

PyTorch Triton Distributed training

Honors

International Mathematical Olympiad

Honorable Mention, 2025 · Best performer in Macau across MO selection tests, IMO, China MO, and China TST

Asian Physics Olympiad

Bronze Medal, 2024 · Qualified for IPhO 2024

International Olympiad in Artificial Intelligence

Bronze Medal, 2024 · ML, computer vision, and NLP problem solving

Chinese Mathematical Olympiad

Bronze Medal, 2025

Chinese Physics Olympiad

Bronze Medal ×2, 2023 & 2024

Van Amringe Mathematical Prize

Columbia Math Prize Exam, 2026 · Single freshman recipient

Columbia Integration Bee

Grand Integrator, 2026 · First freshman champion in the competition's history

Yau MathCamp

Best Project Award, 2024 · Research on the Inverse Galois Problem · View project

Skills

Machine LearningPythonPyTorchC++ Hugging FaceTritonQuantum Computing Math ModelingProbability & StatisticsMATLAB LaTeXGame TheoryPhysics

Outside of coursework, I'm interested in transportation systems, especially railways, along with Rubik's cubes and paper models. See some creative works →

Contact

Reach me by email or find me on GitHub.

jhliang@mit.edu github.com/flbjcy