Seunghyun Yoo

PERSONAL WEBSITE

Seunghyun Yoo

Research Fellow @ Supervised Program for Alignment Research

Undergraduate Researcher @ GRAIL

yoo236@purdue.edu

seunghyun050616@gmail.com

About

Track my TFT rank (for fun)

Ranked · NA

Loading rank…

Hi, I'm Seunghyun! I'm an AI undergrad at Purdue (minoring in math), a Research Fellow at the Supervised Program for Alignment Research, and an Undergraduate Researcher at the Governance and Responsible AI Lab (GRAIL). What I really want is to do research that truly helps humanity. For me, that means finding the interpretability technique or model architecture that actually fits a problem, then using it to solve the problem outright, figure out what's really going on, or lay the groundwork for whoever cracks it next (maybe me!).

When I'm not staring at activations, I'm probably eating pizza (a lot of it, honestly) or playing Overwatch and TFT, and I'm always up for chatting about any of them!

InterestsMechanistic InterpretabilityTechnical AI SafetyMulti-Agent Systems

Education

Purdue University

Aug. 2024 - Expected Dec. 2027

B.S. in Artificial Intelligence

Minor in Mathematics · West Lafayette, IN

Publications

Workshop Papers

Animated architecture diagram: a policy passage enters Gemma 2 2B, drawn as a residual stream with attention and MLP blocks. At layer l, a logit lens, a sparse autoencoder and a steering vector read from or write to the stream. The resulting evidence package goes to a fixed auditor whose report is reviewed: the hybrid interface scores most useful, and a report citing shuffled evidence is flagged.
White-box tools on the residual stream feed an evidence package for a fixed auditor.

White Box Evidence Packages for Policy Audit Reports

Seunghyun Yoo

Second Workshop on Technical AI Governance Research (TAIGR), ICML 2026

PaperPoster
Animated architecture diagram: a diffusion language model with one residual stream per position and bidirectional attention between positions. Each denoising step commits confident tokens and feeds them back. An SAE reads layer l, and a heatmap tracks mask-tracking, commitment and unmasked-token features across 64 steps.
SAE features across 64 denoising steps: mask tracking, commitment, unmasked responses.

Feature Evolution in Diffusion Language Models

Meliksah Yorulmazlar, Justice Sefas, Seunghyun Yoo, and Lulu Shao

Non-Autoregressive Language Models Workshop, COLM 2026 · Poster

PaperPosterLive demo

Projects

Clip from the Ethogram demo: the behaviour atlas plays through the record's phases, with behaviour clusters, influence links lighting up, and a phase timeline along the bottom.
Open the live behaviour atlas.

AI Village × Grove Research AI Swarm Dynamics Hackathon · Oct. 2026

Ethogram

An evidence-linked behaviour catalogue and interactive atlas for multi-agent logs. Tracks how behaviours change and spread across AI agents, with every claim linked to the events behind it.

Live demo Code
Animated architecture diagram: inside the arena an agent LLM, drawn as a residual stream with attention and MLP blocks, generates tokens, calls a sandbox and a verifier, and draws down five budgets. Activation hooks cache internal states; when the token budget runs out the trial ends and the saved record feeds SAE and probe analysis.
Agent loop under five budgets, with cached activations.

AI Incident Response Sprint · Apart Research · Sep. 2026

Top 25% of 230+ projects

MARGIN Arena: Agents Under Resource Limits

A reproducible environment for studying AI agents under token, time, verification, tool-call, and session limits. Connects external tasks with inspectable model records.

Paper Code
Animated architecture diagram: LLaDA drawn as a residual stream over denoising steps. An SAE at layer l feeds a monitor score whose trace crosses the threshold at step 9.6 and triggers steering into the stream; an output-only monitor detects only at step 63.
SAE trajectory monitor flags at step 9.6; output-only at 63.

AI Control Hackathon · Apart Research / Redwood Research · Mar. 2026

Denoising Trajectories as Early-Warning Signals for Diffusion Language Model Control

SAE-based monitoring and steering for diffusion language models. Trajectory monitoring detects attacks at a mean denoising step of 9.60, compared with 63.00 for output-only checks on the main benchmark.

Paper Code

Teaching

AI study session at OZ Lab: students working together on laptops around classroom tables.

Sep. 2025 - Present

LikeLion US / OZ Lab

AI Study Leader

Teach LLM From Scratch, MiniArena, and Interpretability for Purdue's Korean student community.

Honors

Full Member, Sigma Xi

Jul. 2026

The Scientific Research Honor Society

Travel Grant, Grove Research

2026

$1,500

Travel Grant, TAIGR Workshop at ICML 2026

2026

$1,000

Travel Grant, RISE AI Conference

Oct. 2025

University of Notre Dame

Vitæ

Download CV
  • Supervised Program for Alignment ResearchSep. 2026 - Present
    Research Fellow
    Mentors: Yuxiao Li and Di Wu · Remote
  • Governance and Responsible AI Lab (GRAIL)Jun. 2025 - Present
    Undergraduate Researcher
    Purdue University
  • Algoverse AI Safety FellowshipFeb. 2026 - Jun. 2026
    Research Fellow
    Supervisor: Justice Sefas · Remote
  • Purdue UniversityAug. 2024 - Expected Dec. 2027
    B.S. in Artificial Intelligence
    Minor in Mathematics · West Lafayette, IN