Tony Gao

A little about me.

I’m a student at Horace Mann, class of 2027, and most of my time outside class goes to machine-learning research. I'm interested in mechanistic interpretability and AI safety.

These are the projects I’ve spent the most time on. Some I did on my own and some with coauthors, and I’ve tried to say which.

Papers
NeurIPS 2026 workshop (accepted) · arXiv preprint 2026 · NHSJS 2024
Coursework
Carnegie Mellon, summer 2026: Matrices and Linear Transformations
Tools
PyTorch, TransformerLens, sparse autoencoders, linear probes, activation steering, PPO

Things I’ve
worked on.

Do features that predict an agent’s behavior also control it?

When Probes Fail to Steer is my own project, and a NeurIPS 2026 workshop paper. I trained a 1,024-feature sparse autoencoder on a frozen Rocket League agent, along with probes that predict its behavior at 0.89 to 0.96 AUROC, then tried 624 edits to its activations while it played.

Most of the edits didn’t change what the agent did. One SAE feature raised the dribble rate from 42.8 to 65.1 per 1k steps, on 11 of 12 held-out seeds.

Dribbles per 1k steps, from 42.8
65.1
Held-out seeds
11 / 12
KL from the unedited policy
0.0047

How often do language models write insecure cryptography code?

CIPHER is a benchmark I built with a co-first author: 450 prompts written three ways (insecure, neutral and secure) and a taxonomy of 58 vulnerability types. We ran it on seven LLMs and scored the code with pattern detectors and an LLM judge that points to the lines it flags.

Between 56% and 89% of the generated code had flaws. Asking for secure code only lowered the overall rate from 76.9% to 71.7%.

Prompts, asked three ways
450
LLMs evaluated
7
Still flawed with a security prompt
71.7%

A trading competition with my team.

IMC Prosperity 4 is a 16-day trading simulation with algorithmic and manual rounds on market making, options and risk. I led our team and wrote our trading algorithms in Python. We finished in the top 5% of 18,803 teams.

Worldwide
Top 5%
Teams
18,803
Days
16

Forecasting stock returns and volatility.

A personal project, since 2025, where I compare gradient boosting, Gaussian processes, LSTMs and hidden Markov models, and train a PPO agent to split money across several assets.

Most of the work went into the backtester. It splits data by date and checks for leakage, so a model is never tested on data it has already seen.

Model families
4
Multi-asset allocator
PPO
Train and test splits
By date

Rating YouTube tutorials from their comments.

My first paper, which I co-authored, published in NHSJS in 2024. We rated 50 tutorials using engagement and the sentiment of their comments. Comment sentiment was the best single predictor, and a random forest did better than the six other models we tried.

r, sentiment vs rating
0.85
Test MSE, vs 0.195 linear
0.06
Videos
50

Other things I do.

I’ve played piano for over ten years, and performed multiple times at Carnegie Hall’s Weill Recital Hall.

I’m also president of program planning at Citizens of Tomorrow, which offers free English classes to immigrant learners. I co-lead 15 classes for about 200 learners and help coordinate about 100 student volunteers.

Honors
National Merit Semifinalist · USACO Silver · President’s Volunteer Service Award