Andrew Sung

systems, models, and the spaces between them
Andrew Sung

Andrew Sung

drewsung@stanford.edu
github linkedin twitter dogs

I'm a senior at Stanford studying computer science (AI track), admitted to the coterminal MS program. I'm drawn to hard open problems — making AI systems more reliable and easier to trust.

That thread runs through my work: test-time training when verification is binary (TTT-Binary), tool-grounded agents that look things up instead of hallucinating (MCPkmn Showdown), RL for structured decision-making under uncertainty (Wordle), and translating mycelium electrical signals into sound and movement (A Thousand Hyphae).

Previously: ByteDance (Global Payments Privacy & Security), Sorano AI (cofounder, TD3-based weather forecasting), Stanford HCI Group & AI Lab (memory systems for safe long-horizon LLM coaching). Currently doing ML research at Stanford on representation learning.

Coursework: Frontier Systems · AI Agents & Simulations · RL · NLP · Mechanistic Interpretability · Decision Making Under Uncertainty · Mechatronics

scroll
Selected work
01

TTT-Binary

Improving LLM reasoning reliability at inference time. GRPO on self-generated curricula pushed FrontierMath solve rate from 0.67% to 11.80%.
researchGRPOtest-time training
02

MCPkmn Showdown

Open-source MCP to reduce LLM hallucination in Pokémon Showdown. 1st place / 800 @ Stanford's Inaugural Claude Builder Challenge.
open-sourceLLM agentshallucination reduction
03

RL Wordle Agent

Entropy-guided PPO for structured decision-making. Information-theoretic priors + policy gradients achieve 100% win rate in 3.51 avg guesses.
researchPPOreward shaping
04

A Thousand Hyphae

Mycelium electrophysiology signals mapped to sound and physical movement. 7-electrode spike-triggered audio, motorized rotating sculptures.
installationhardwaresonification
Technical Interests RL · LLMs · Mech Interp · Test-Time Training · Agents · Simulations
Nontechnical Interests Chess · Golf · Boxing · Cello · Scuba · Fermentation · Astronomy
About
Andrew Sung
Andrew Sung
Stanford CS '26 · MS CS (admitted)
I'm drawn to hard open problems, like making AI systems more reliable and easier to trust. My work lives at the intersection of learning, decision-making, and systems that work in the real world. Sometimes as a research paper, sometimes as an art installation.
Currently: ML Research, Stanford
Previously: ByteDance · Stanford HCI Group & AI Lab · Sorano AI

Coursework: Frontier Systems · AI Agents & Simulations · RL · NLP · Mechanistic Interpretability · Decision Making Under Uncertainty · Mechatronics
GitHub LinkedIn Twitter
somewhere in the loop, the surface twists