Jasmine Shone

Student + Researcher at MIT

Hi! I'm Jasmine Shone, a current student at MIT. This summer, I was a SWE Intern at Meta working on a computer vision model launch at Meta Superintelligence Labs and large-scale load testing at Instagram.

I am currently researching at the Kaiming He Lab, focusing on effective tokenization/representations for large scale chaotic system data in combination with latent diffusion models. I'm one of the co-leads for AI @ MIT Reading Group (and have been in the group for a total of 3 semesters now).

Jasmine Shone

Research

Latent Diffusion

Latent Diffusion for PDEs

In Progress

Modeling cool physics

Contributions
Building out pipeline, running experiments

SAM 3D @ Meta Superintelligence

In Submission

2D to 3D objects!

launch paper
Contributions
Designing and implementing a modified version of the SuGaR algorithm, reducing runtime from 30 m to ~1 m 40 s. Proposing and implementing pointcloud downsampling algorithm which reduces points by up to an 12% average. Created parallelized script to improve training efficiency through data cleaning, detecting 6+% of corrupted data. Implemented baseline for evaluation metric and created easy-to-follow jupyter notebook.
Beyond I-Con

Beyond I-Con: Exploring a New Dimension of Distance Measures

NeurIPS NeurReps Workshop 2025 · NeurIPS WiML Workshop 2025

We extend the I-Con framework to discover new losses which achieve state-of-the-art results on clustering and supervised contrastive learning.

arXiv Blog
Contributions
Designed, implemented, and evaluated loss functions based on the I-Con framework. Conducted analysis on experimental results.
Embeddings

Text-Invariant Similarity Learning

We create a new paradigm of image-pair similarity learning conditioned on text descriptions.

Blog
Contributions
Designed and implemented image-pair similarity annotation pipeline. Designed, implemented, and explored different experimental setups and loss functions.

Keypoint Abstraction using Large Models for Object-Relative Imitation Learning

ICRA 2025 · CoRL LangRob Best Paper 2024

Utilizing priors from Vision-language models and image features to generalize effectively across object poses, camera viewpoints, and object instances with only 10 demonstrations.

arXiv Project
Contributions
Trained multimodal models utilizing pointcloud encoders, object-wise transformers, and vision foundation models, improving evaluation performance by 27.5%
Created and designed keypoint proposal pipeline with specialized VLM prompting SAM, furthest point sampling, mark-based prompting, RANSAC, and point cloud clustering
Datathon

The Correlation of Regional Gas Prices with Unemployment

Citadel Women's Datathon Winning Report 2024

Are fluctuations in gas prices predictive of unemployment rates and how do regional differences in mass transit and gasoline production affect this trend?

Contributions
Coming up with the initial research question + insight to use multiple additional datasets beyond the ones we were provided in the competition. Also worked on most of the stats/modelling work (Granger Causality, ADF, VAR modelling).

SketchAgent: Language-Driven Sequential Sketch Generation

CVPR 2025

We introduce SketchAgent, a novel framework for generating sequential sketches from language prompts.

Project
Contributions
Worked on moving the LLM inference pipeline to a locally hosted version of Llama Vision 3.2 11B and Llama 3.2 90B. Learned about quantization, multi-gpu inference, and prompting techniques.
App Creation

Systematic Optimization of App Generation Few Shot Learning

RSI 2022 · Regeneron STS Top 300

We create an LLM in-context learning pipeline to systematically optimize prompt token length, few-shot selection, and ordering.

Paper
Contributions
Developed the p-mRMR algorithm and prompt creation framework, designed and wrote an evaluation suite of test applications and performed manual error evaluation of generated apps.
Fundus

Stable Diffusion on Glaucoma Fundus Images in Low-Data Settings

Computer Vision Final Project

We improve stable diffusion's ability to generate high-quality fundus images by finetuning on an extremely small dataset of 170 images.

Contributions
Worked on designing the project/experiments, training the model, doing evals with resnet, and baselines (GANs)

Experience

Meta / SWE Intern

Meta Superintelligence Labs (Vision Model Launch) & Instagram (Load Testing).

Hudson River Trading / Core SWE & Algo Intern

Built C++ infra for live trading; Deployed profitable trading bot for Brazilian equities.

MIT LIS Lab / Research Assistant

Robotics generalization research using keypoint abstraction.

Awards

Fun Stuff

Music

Until I Found You (Cover) After the Journey (Orig) Instagram

Writing

Thoughts in the Silent Night Houston Rain

Games

The Collector A Quarky Tale
© 2025 Jasmine Shone.