CV

CV of Ryan Kim — ML research at MedARC, Memorial Sloan Kettering, and Weill Cornell Medicine; CS at UIUC.

Contact Information

Name Ryan Kim
Professional Title CS Undergrad, UIUC
Email ryankim17920@gmail.com
Phone 224-517-7069

Professional Summary

First-author paper at a NeurIPS 2026 workshop Sole-author ICLR 2027 submission Founding team, a16z AI startup

Experience

  • 2026 - Present

    Remote

    RL Intern (Founding Team)
    Delightful AI (a16z speedrun-backed)
    • Built bioinformatics agent RL environments (Chimera) end to end that hold GPT-6 Astra, Claude Opus 5.5, and Fable 5.1 to <0.5/1.0 reward: a synthetic generator of verified-solvable tasks, Docker/Harbor execution, and reward grading.
    • Sped up task generation 6x by adding caching and parallelism to the generator pipeline.
    • Analyzed agent trajectories with an LLM-as-a-judge that checks each run against expected milestones.
  • 2026 - Present

    Urbana, IL

    Undergraduate Researcher
    BLENDER Lab, UIUC (with Alexi Gladstone)
    • Led research extending Energy-Based Transformers to multi-token prediction through Explorative Modeling (XM).
    • Implemented the full training stack and compared candidate energy formulations against block-diffusion LLMs.
  • 2026 - Present

    Remote

    ML Research Contributor
    MedARC
    • Drove 37% of the gain over baseline on NanoPath v1 (26% on v2) by building an LLM autoresearch agent with dual-branch search and anti-overfitting cycles (3rd-author paper, accepted at a NeurIPS 2026 workshop).
    • Cut config errors 80% and token use 15x by designing agent infrastructure (knowledge graphs, SLURM, worktree isolation).
    • Raised mean PathoROB robustness 0.726 to 0.891 and improved downstream AUC across 7 pathology foundation models by introducing SPECTRA, label-free contrastive LoRA fine-tuning (first-author paper, accepted at a NeurIPS 2026 workshop).
  • 2024 - 2026

    New York, NY

    Pathology AI Research Intern
    AI in Medicine & Computational Biology Lab, Weill Cornell Medicine
    • Found that error doubles out of distribution (MAE ~0.10 to ~0.20) by benchmarking 12 vision foundation models for tumor-purity regression on TCGA and internal cohorts in PyTorch.
  • 2024 - 2026

    New York, NY

    Computational Oncology Research Intern
    Advanced Computing & Oncology Lab, Memorial Sloan Kettering
    • Engineered a CPU-only LLM pipeline that classified 370 clinically annotated lung-cancer CT reports, optimizing prompts for Llama 3.1-8B and Qwen3-4B with DSPy MIPROv2 (3rd-author abstract, accepted at ACRO 2026).
    • Applied Cox regression and survival analysis to 230 patients’ clinical and demographic records to identify correlations between percent thymic tissue and NSCLC outcomes.
  • 2025 - 2026

    Stony Brook, NY

    Simons Summer Research Fellow (Computational Genomics)
    Yurovsky Lab, Stony Brook University (Simons Foundation)
    • Predicted survival of 185 glioblastoma patients with a DNABERT-based Cox model on non-coding DNA mutations.
    • Identified significant age-related survival signals and racial disparities in mutational scores.

Education

  • 2026 - 2029

    Urbana, IL

    B.S.
    University of Illinois Urbana-Champaign
    Computer Science
    • Expected May 2029
    • Junior Standing
    • GPA: 4.0/4.0
    • Chancellor’s Scholar
    • AIME Qualifier (4x)

Projects

  • Which State Should Prediction Read?

    August 2026 - September 2026 Sole-author paper, under review at ICLR 2027
    • Trained State–Prediction Separation (SPS) LMs, which split a Transformer into separate state and prediction streams; ablations showed the two divide the work: state tracks context, prediction retrieves from it.
    • Reached a FLOP-matched Transformer’s loss on 2.1x fewer tokens and ended 0.064 nats/token lower after 20B tokens by designing Sequential, a final-state-read architecture, from these ablations (92M non-embedding params).
  • Qwen3.6-27B ExpThink

    July 2026 - August 2026 LLM Post-Training, GRPO
    • Cut thinking tokens 52% in-domain / 34% OOD and truncation 6.0% to 2.6% while mean OOD accuracy rose 84% to 86% by post-training Qwen3.6-27B with a length-adjusted GRPO reward.
    • Trained with PyTorch FSDP on 8xH100 in verl, with rollouts from tensor-parallel SGLang.
  • speedLM

    July 2026 - August 2026 OpenAI-Compatible LLM Serving, Speculative Decoding
    • Built a custom vLLM server that fine-tunes EAGLE-3 draft models on captured local traffic during idle GPU time.
    • Raised tokens accepted per verification step 13.4% over the stock draft on held-out Qwen3-8B data.
  • micro-DINOv3

    February 2026 - April 2026 Pure Python, Computer Vision
    • Implemented DINOv3 in dependency-free Python with hand-written reverse-mode autograd and a ViT, EMA teacher-student distillation, DINO + iBOT objectives, KoLeo regularization, axial RoPE with register tokens, and Gram anchoring.
  • DeepMoE Reproduction

    July 2024 - November 2024 PyTorch, Computer Vision
    • Recreated the main CIFAR-10 results of “Deep Mixture of Experts via Shallow Embedding” (100+ citations) in PyTorch, recovering its accuracy-vs-compute tradeoff.
    • Identified a gate-scaling failure that destabilizes training at the paper’s learning rate.

Skills

LLM Stack: PyTorch, FSDP, verl, SGLang, vLLM, Hugging Face, DSPy, EAGLE-3
Machine Learning: Large Language Models (LLMs), Deep Learning, Transformers, Reinforcement Learning, Distributed Training
Methods: LLM Architecture & Interpretability, Post-Training (GRPO), Speculative Decoding, RL Environments, LLM Agents
Languages: Python, C/C++, Java, SQL (PostgreSQL), JavaScript (React, Node.js)
Tools: SLURM, Docker, Harbor, Linux, Git, FastAPI, Scikit-learn, Polars, Pandas, NumPy

Publications

  1. Which State Should Prediction Read? A Mechanistic Analysis of State–Prediction Separation

    Ryan Kim

    ICLR 2027 Under review

    Under review at ICLR 2027

    Code

  2. From Text to Treatment: Automated Identification of Airway Events and Referral Patterns in Lung Cancer Imaging Using Locally Deployed CPU-Only LLMs

    Kwadwo Amoako-Boadu, Alexander Kim, Ryan Kim, David Miller, Rami Naaman, Abraham J. Book, Robert Lee, Narek Shaverdian, Daniel Gomez, Tafadzwa L. Chaunzwa

    ACRO 2026 Annual Meeting Accepted abstract

  3. Contrastive fine-tuning against stain and scanner bias improves robustness and downstream performance across pathology foundation models

    Ryan Kim, Hassan Muhammad, Anish Dulal, Paul Steven Scotti, Tanishq Mathew Abraham

    NeurIPS 2026 Workshop Accepted

    NeurIPS 2026 Workshop on AI at Scale for Clinical Impact: Cancer Pathology Foundation Models (ASCI)

    OpenReview Code

  4. NanoPath: Fast, Fair, Open Experimentation for Pathology Foundation Models

    Paul Steven Scotti, Hassan Muhammad, Ryan Kim, Nima Ashjaee, Ratna Sagari Grandhi, Saman Rahbar, Raymond Biju, Anish Dulal, Chao Feng, Connor Lane, Benjamin Warner, Tanishq Mathew Abraham

    NeurIPS 2026 Workshop Accepted

    NeurIPS 2026 Workshop on AI at Scale for Clinical Impact: Cancer Pathology Foundation Models (ASCI)

    OpenReview Code Website