// Shared CV data, sourced from Kang_Minjae_CV.pdf + Google Scholar.

const PROFILE = {
  name: 'Minjae Kang',
  nameKo: '강민재',
  role: 'LLM Researcher',
  affiliation: 'Machine Learning and Language Lab (ML3)',
  university: 'Yonsei University',
  advisor: 'Jaehyung Kim',
  advisorUrl: 'https://sites.google.com/view/jaehyungkim',
  labUrl: 'https://ml3.yonsei.ac.kr/',
  location: 'Seoul, South Korea',
  emails: ['mjkang618@yonsei.ac.kr', 'kminjae618@gmail.com'],
  links: {
    scholar: 'https://scholar.google.com/citations?user=OpeDkEkAAAAJ&hl=ko',
    github: 'https://github.com/mjk0618',
    linkedin: 'https://www.linkedin.com/in/%EB%AF%BC%EC%9E%AC-%EA%B0%95-959888273/',
  },
};

const BIO_FIRST = [
  "Hi, I'm at the Machine Learning & Language Lab (ML3), Yonsei University, advised by Prof. Jaehyung Kim.",
  "My research centers on alignment and safety for LLMs: making language models more controllable, reliable, faithful to user intent, and easier to understand well enough to intervene on.",
  "Recently I've been working on activation steering, using directions in hidden representations to improve instruction following at inference time without retraining.",
];

const RESEARCH_INTERESTS = [
  {
    title: 'Alignment & Safety',
    body: 'Building LLMs that behave the way we intend, even under distribution shift, adversarial pressure, or ambiguous instructions. I am interested in controllability, faithful instruction following, and interpretability as practical tools for making model behavior more reliable.',
  },
  {
    title: 'Efficient LLMs & Optimization',
    body: 'Lowering the barriers to using LLMs by reducing deployment cost, compute requirements, and access constraints. I am interested in efficiency-oriented methods that make capable language models more broadly usable for researchers, developers, and everyday users.',
  },
];

const NEWS = [
  { date: 'Jul 2026', tag: 'travel', text: 'Will be attending **ICML 2026** in Seoul, South Korea.' },
  { date: 'Apr 2026', tag: 'travel', text: 'Will be attending **ICLR 2026** in Rio de Janeiro, Brazil.' },
];

const PUBS = [
  {
    id: 'directer',
    title: 'Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection',
    short: 'DIRECTER',
    authors: ['Minjae Kang', 'Jaehyung Kim'],
    venue: 'ICLR',
    year: 2026,
    status: 'Accepted',
    topics: ['activation steering', 'instruction following'],
    abstract: 'A test-time activation-steering method that improves instruction-following by dynamically rejecting candidate generations whose internal states drift from a learned compliance direction. Strong gains on IFEval-style benchmarks without any retraining.',
    links: [{ label: 'arXiv', url: 'https://arxiv.org/abs/2603.06745' }],
  },
  {
    id: 'cope',
    title: 'Personalized LLM Decoding via Contrasting Personal Preference',
    authors: ['Hyungjune Bu', 'Chanjoo Jung', 'Minjae Kang', 'Jaehyung Kim'],
    venue: 'EMNLP',
    year: 2025,
    status: 'Main',
    topics: ['personalization', 'instruction tuning'],
    abstract: 'Optimizes the in-context prompt per-user to elicit personalized behavior from instruction-tuned models without changing weights. Bridges in-context learning and lightweight personalization.',
    links: [{ label: 'arXiv', url: 'https://arxiv.org/abs/2506.12109' }],
  },
  {
    id: 'rstiefel',
    title: 'Riemannian Optimization for LoRA on the Stiefel Manifold',
    authors: ['JuneYoung Park', 'Minjae Kang', 'Seongbae Lee', 'Haegang Lee', 'Seongwan Kim', 'Jaeho Lee'],
    venue: 'EMNLP Findings',
    year: 2025,
    status: 'Findings',
    topics: ['LoRA', 'optimization'],
    abstract: 'Constrains LoRA factors to the Stiefel manifold and trains with Riemannian optimization. Improves LoRA stability and downstream performance at matched parameter counts.',
    links: [{ label: 'arXiv', url: 'https://arxiv.org/abs/2508.17901' }],
  },
];

const EXPERIENCE = [
  {
    org: 'Yonsei University · ML3 Lab',
    role: 'Graduate Researcher',
    when: 'Apr 2025 – Feb 2026',
    bullets: [
      'Activation engineering and inference-time control for instruction-tuned LLMs.',
      'Advised by Prof. Jaehyung Kim.',
    ],
  },
  {
    org: 'Opt-AI Inc.',
    role: 'LLM Researcher (Quantization)',
    when: 'Oct 2024 – May 2025',
    summary: 'On-device LLM research, domain-specific adaptation, and quantization.',
    bullets: [
      'Deployed LLMs to NPU environments using the executorch framework while preserving task performance.',
      'Applied PTQ algorithms (e.g., SpinQuant) and ran calibration experiments to minimize quantization loss.',
    ],
  },
  {
    org: 'NAVER · AX Center',
    role: 'LLM Research Intern',
    when: 'Jan 2024 – Feb 2024',
    summary: 'Agentic workflow for review analysis using LLM-based clustering.',
    bullets: [
      'LLM-based clustering on NAVER review data without separate embedding models.',
      'Proposed an Iterative DPO method using GPT-4-generated synthetic data as ground truth.',
      'Fine-tuned HCX models for clustering, beating embedding-based baselines.',
    ],
  },
  {
    org: 'BNT Solution',
    role: 'Big Data Research Intern',
    when: 'Jul 2022 – Aug 2022',
    summary: 'Sea-surface-temperature pipeline for a big-data platform.',
    bullets: [
      'Built a processing pipeline for sea-surface-temperature data.',
      'Contributed to a government-led big data platform project (NIA).',
    ],
  },
];

const EDUCATION = [
  {
    school: 'Yonsei University',
    degree: 'M.S. Student, ML3 Lab',
    when: 'Mar 2026 – Present',
    note: 'Advised by Prof. Jaehyung Kim',
  },
  {
    school: 'Sungkyunkwan University',
    degree: 'B.S., Electronic and Electrical Engineering',
    when: 'Mar 2018 – Feb 2025',
  },
];

const PROJECTS = [
  {
    title: 'LawBot — Legal Advisory Chatbot',
    when: 'Jun 2023 – Aug 2023',
    body: 'A LLaMA-2-based legal chatbot for dispute resolution. Built end-to-end automation for data generation, collection, and preprocessing; designed quantitative eval metrics over output embeddings.',
    link: 'https://github.com/mjk0618',
  },
];

const TALKS = [];

const TEACHING = [];

const ACADEMIC_SERVICES = [
  {
    category: 'Conference Reviewer',
    items: [
      {
        title: 'Neural Information Processing Systems (NeurIPS)',
        years: '2026',
        url: 'https://neurips.cc/',
      },
    ],
  },
];

Object.assign(window, {
  PROFILE, BIO_FIRST, RESEARCH_INTERESTS, NEWS, PUBS, EXPERIENCE, EDUCATION, PROJECTS, TALKS, TEACHING, ACADEMIC_SERVICES,
});