Chris Ge

Hi, I’m Chris Ge. I’ve recently finished my Bachelor’s in CS (course 6-3) at MIT, and I’m now pursuing a 4th year MEng in CS. My research interest is in using interpretability to help understand and control the models people actually use, particularly for AI safety applications.

At MIT, I’m a part of Antonio Torralba Lab, where I do mechanistic interpretability research on models of practical interest, including multimodal diffusion transformer models like FLUX.2. I also have industry experience with agents: at Fulcrum, I trained models to predict how likely an agent is to succeed on a given task, and at Hudson River Trading (research intern on HRT AI Labs), I worked on AI agent autoresearch setups for model development. Outside the lab, I’m a TA for Introduction to Machine Learning (6.3900) at MIT, where I teach a recitation section and help develop course content.

publications

Vision-Language Binding in In-Context Image Generation

Chris Ge, Rohit Gandikota, Antonio Torralba, and Tamar Rott Shaham

Mechanistic Interpretability Workshop @ ICML 2026 (Spotlight); Under review

arXiv preprint arXiv:2605.24624, 2026

arXiv Bib Code Poster Website

@article{ge2026vision,
  title = {Vision-Language Binding in In-Context Image Generation},
  author = {Ge, Chris and Gandikota, Rohit and Torralba, Antonio and Rott Shaham, Tamar},
  journal = {arXiv preprint arXiv:2605.24624},
  year = {2026},
  venue = {Mechanistic Interpretability Workshop @ ICML 2026 (Spotlight); Under review},
}

Agent psychometrics: Task-level performance prediction in agentic coding benchmarks

Chris Ge^*, Daria Kryvosheieva^*, Daniel Fried, Uzay Girit, and Kaivalya Hariharan

COLM 2026

arXiv preprint arXiv:2604.00594, 2026

arXiv Bib Code

@article{ge2026agent,
  title = {Agent psychometrics: Task-level performance prediction in agentic coding benchmarks},
  author = {Ge, Chris and Kryvosheieva, Daria and Fried, Daniel and Girit, Uzay and Hariharan, Kaivalya},
  journal = {arXiv preprint arXiv:2604.00594},
  year = {2026},
  venue = {COLM 2026},
}

Private Linear Regression via a Down-Sensitivity to Privacy Reduction

Ittai Rubinstein, Chris Ge, and Samuel B Hopkins

COLT 2026

Bib PDF

@article{rubinsteinprivate,
  title = {Private Linear Regression via a Down-Sensitivity to Privacy Reduction},
  author = {Rubinstein, Ittai and Ge, Chris and Hopkins, Samuel B},
  venue = {COLT 2026},
}

other projects

Team Leduc Poker

We used multi-agent learning algorithms to find a correlated equilibrium of the Team Leduc Hold'em game, a variant of Leduc poker where players 1 & 3 and players 2 & 4 play as a team without being able to communicate privately during the game. We submitted our strategies as part of the poker competition for MIT 6.S890 Multiagent Learning and got 2nd place.

Code Report Presentation

Phoneme Bridges

We augment textual data with phonetic information in order to improve cross-lingual transfer from Hindi, a high-resource language, to Urdu, a low-resource language that shares many words but with a different written script. Done as a group project for 6.8611 Natural Language Processing.

Code Report Poster