About me

I am a Ph.D. candidate in the Department of Electrical Engineering and Computer Science (EECS) at MIT, where I am fortunate to be advised by Prof. Julian Shun. I received my M.S. in Computer Science at MIT, and my B.S. in Electronic and Information Engineering from Huazhong University of Science and Technology, in 2019.

My research focuses on Machine Learning and Data Mining, with an emphasis on developing effective and efficient algorithms to tackle high-impact, real-world problems. I am particularly interested in large language models, graph neural networks, and scalable learning algorithms. I have collaborated with Prof. Dawei Zhou at Virgina Tech on reasoning with large language models, and Dr. Yada Zhu at IBM Research IBM Research on applying graph learning to financial fraud detection.

What I'm Focusing

  • efficient icon

    Efficient ML Algorithms

    Designing efficient and scalable machine learning algorithms to tackle high-impact real-world problems with performance and reliability.

  • enhancement icon

    LLM Enhancement

    Developing advanced techniques to improve the reasoning, adaptability, and efficiency of large language models across diverse domains.

  • quant icon

    Quantitative Research

    Applying statistical and machine learning methods to quantitative modeling, risk analysis, and systematic trading strategies.

  • codesign icon

    ML + System Co-Design

    Exploring the intersection of machine learning and systems, co-designing algorithms with innovative hardware and software to enhance data acquisition, processing, and real-world analytics.

Latest News

Resume

Education

  1. Massachusetts Institute of Technology

    2022 – 2026 (expected)

    Department of Electrical Engineering & Computer Science (EECS)
    Computer Science & Artificial Intelligence Laboratory (CSAIL)
    PhD Candidate in Computer Science

  2. Massachusetts Institute of Technology

    2019 – 2022

    Department of Electrical Engineering & Computer Science (EECS)
    MIT Research Laboratory of Electronics
    Master of Science in Computer Science

  3. Huazhong University of Science and Technology

    2015 – 2019

    School of Electronic Information and Communications (EIC)
    Bachelor of Engineering in Electronic and Information Engineering (EE)

Experience

  1. Quantitative Research Intern

    May 2025 — Present

    Quantbot Technologies LP, New York, USA

  2. Graduate Research Assistant - Understanding Underlying Mechanisms of Large Language Models

    Jan 2025 — Present

    MIT CSAIL (Prof. Julian Shun), Cambridge, Massachusetts, USA
    • Developed Plan-and-Budget, a test-time scaling framework (~3,000 lines of Python) that improves LLM reasoning efficiency by dynamically planning subquestions and budgeting tokens, yielding up to +70% accuracy, -39% token reduction, and +187.5% improvement in ℰ³ metric.
    • Demonstrated Plan-and-Budget’s capability to close model efficiency gaps without retraining, enabling DS-Qwen-32B to match the efficiency of DS-LLaMA-70B across reasoning benchmarks.
    • Co-developed LensLLM, an interpretability and model selection framework that analyzes fine-tuning dynamics across checkpoints, achieving up to 91.1% accuracy and reducing LLM selection cost by 88.5%, outperforming five state-of-the-art baselines.
    • Co-authored two papers accepted to ICML 2025 and under review at top-tier venues.

  3. Machine Learning Student Researcher - Large Language Model Guided Graph Learning

    August 2023 — Present

    Watson AI Lab, IBM Research, Cambridge, Massachusetts, USA
    • Proposed a bidirectional reasoning framework between large language models (LLMs) and knowledge graphs (KGs), enabling KG-augmented LLM inference and LLM-assisted KG evolution.
    • Developed a temporal reasoning framework (~15,000 lines of Python) that significantly outperforms prior methods, achieving up to 23.3% gain in temporal reasoning tasks and 8% gain in evolving KG scenarios.
    • Demonstrated that we bridged the capability gap between model sizes—e.g., a LLaMA 3.1–8B model improved from 18.6% to 37.0%, nearly matching a much larger DeepSeek-V3 671B model (38.3%).
    • Proposed a super-relation reasoning method that aggregates relation paths into semantically meaningful units to enhance KGQA, yielding +2.92% accuracy over baselines across nine real-world datasets.
    • Co-authored two papers accepted to ICLR 2025 and under review at top-tier venues.

  4. Machine Learning Research Intern - Graph Neural Network Enabled Financial Fraud Detection

    June 2023 — August 2023

    Watson AI Lab, IBM Research, Cambridge, Massachusetts, USA
    • Designed the first real-world heterophilic and heterogeneous graph benchmark (ℋ²GB, published in KDD)
    • Designed the first unified, scalable graph transformer framework (UnifiedGT, published in IEEE Big Data) on large-scale graph mining and implemented the framework from scratch using ~3000 lines of Python code.
    • Conducted comprehensive experiments using PyTorch and PyTorch Geometric Library (PyG).
    • Achieved state-of-the-art on 9 large-scale datasets proposed in ℋ2GB and enhanced the node classification accuracy by 5-10%, including RCCD (14 million nodes, 160 million edges).
    • Achieved state-of-the-art on 6 Anti-Money Laundering (AML) datasets (180 million edges) and enhanced the link classification F1 score by 8-18% while delivering 2.4× greater throughput and reduced latency.

  5. Machine Learning Research Intern - Graph Neural Network Enabled Financial Fraud Detection

    June 2023 — August 2023

    Watson AI Lab, IBM Research, Cambridge, Massachusetts, USA
    • Designed the first real-world heterophilic and heterogeneous graph benchmark (ℋ²GB, published in KDD)
    • Designed the first unified, scalable graph transformer framework (UnifiedGT, published in IEEE Big Data) on large-scale graph mining and implemented the framework from scratch using ~3000 lines of Python code.
    • Conducted comprehensive experiments using PyTorch and PyTorch Geometric Library (PyG).
    • Achieved state-of-the-art on 9 large-scale datasets proposed in ℋ2GB and enhanced the node classification accuracy by 5-10%, including RCCD (14 million nodes, 160 million edges).
    • Achieved state-of-the-art on 6 Anti-Money Laundering (AML) datasets (180 million edges) and enhanced the link classification F1 score by 8-18% while delivering 2.4× greater throughput and reduced latency.

Publications

Publications

  1. Plan and Budget: Effective and Efficient Test-Time Scaling on Large Language Model Reasoning

    2025

    Junhong Lin, Xinyue Zeng, Jie Zhu, Song Wang, Julian Shun, Jun Wu, Dawei Zhou
    arXiv preprint arXiv:2505.16122.

  2. When Heterophily Meets Heterogeneity: New Graph Benchmarks and Effective Methods

    2025

    Junhong Lin, Xiaojie Guo, Shuaicheng Zhang, Yada Zhu, Julian Shun
    ACM SIGKDD 2025, 8.75/10 High Review Rating.

  3. LENSLLM: Unveiling Fine-Tuning Dynamics for LLM Selection

    2025

    Xinyue Zeng, Haohui Wang, Junhong Lin, Jun Wu, Tyler Cody, Dawei Zhou
    International Conference on Machine Learning (ICML 2025).

  4. Reasoning of Large Language Models over Knowledge Graphs with Super-Relations

    2025

    Song Wang, Junhong Lin, Xiaojie Guo, Julian Shun, Jundong Li, Yada Zhu
    International Conference on Learning Representations (ICLR 2025).

  5. FraudGT: A Simple, Effective, and Efficient Graph Transformer for Financial Fraud Detection

    2024

    Junhong Lin, Xiaojie Guo, Yada Zhu, Samuel Mitchell, Erik Altman, Julian Shun
    International Conference on AI in Finance (ICAIF 2024), Best Paper – Honorary Mention

  6. UnifiedGT: Towards a Universal Framework of Transformers in Large-Scale Graph Learning

    2024

    Junhong Lin, Xiaojie Guo, Shuaicheng Zhang, Dawei Zhou, Yada Zhu, Julian Shun
    IEEE International Conference on Big Data (IEEE Big Data 2024)

  7. Large language models for forecasting and anomaly detection: A systematic literature review

    2024

    Jing Su, Chufeng Jiang, Xin Jin, Yuxin Qiao, Tingsong Xiao, Hongda Ma, Rong Wei, Zhi Jing, Jiajun Xu, Junhong Lin

Patents

  1. Reasoning of Large Language Models Over Knowledge Graphs with Super-Relations

    Pending US Patent

  2. Unified Graph Transformer for Financial Fraud Detection on Massive Graphs

    Pending US Patent

    Yada Zhu, Xiaojie Guo, Junhong Lin, Shuaicheng Zhang, Julian Shun

  3. Indoor Positioning Method based on Weighted Surface Fitting from Crowdsourced Sample

    CN 109059919 B

    Bang Wang, Junhong Lin, Guang Yang

  4. Location Map Building Method based on Virtual Source Estimation and Trajectory Correction

    CN 108919177 B

    Bang Wang, Junhong Lin, Guang Yang

  5. Portable Psychological Pressure Detector

    CN 108294766 A

Portfolio

Awards & Honors

  1. National Science Foundation Student Travel Award

    2024

  2. Best Paper – Honorary Mention at the 5th International Conference on AI in Finance

    2024

  3. 2024 J. Francis Reintjes Excellence in 6A Industrial Practice Award

    2024

    Massachusetts Institute of Technology

  4. Outstanding Undergraduates in Term of Academic Performance in 2017 (top 1%)

    2017

    Huazhong University of Science and Technology

  5. National 1st Prize & 1st Award for Selected Topic in “Renesas Cup" National Undergraduate Electronic Design Contest

    2017

  6. TI Cup Winner & 1st Award in "TI Cup" Hubei Undergraduate Electronic Design Contest

    2017

  7. Meritorious Winner in Mathematical Contest in Modeling, COMAP

    2017

Presentations

  1. UnifiedGT: Towards a Universal Framework of Transformers in Large-Scale Graph Learning

    2024

    IEEE International Conference on Big Data (IEEE Big Data 2024)

  2. FraudGT: A Simple, Effective, and Efficient Graph Transformer for Financial Fraud Detection

    2024

    International Conference on AI in Finance (ICAIF 2024)

Contact

Contact Form