JACK D.V. CARSON

Research

i'm an undergraduate at mit studying electrical engineering with computing (6-5) and mathematics (18). i work as a student researcher in the regina barzilay group at mit csail, and i'm broadly interested in the intersection of deep learning, physics, and the life sciences — stochastic processes, condensed matter theory, graph neural networks, and molecular dynamics.


A Statistical Physics of Language Model Reasoning

READ THE PAPER modeling transformer reasoning as a stochastic dynamical system
also at arXiv:2506.04374 / OpenReview

Transformer language models reason in ways that have largely resisted mechanistic understanding. With Amir Reisizadeh, I introduce a statistical-physics framework that treats chain-of-thought reasoning as a continuous-time process: sentence-level hidden states evolve as a low-dimensional stochastic dynamical system, decomposing each semantic trajectory into deterministic drift and stochastic fluctuation. Across eight open-source models and seven reasoning benchmarks, a rank-40 drift manifold captures roughly 50% of the variance, and the trajectories sort into four distinct latent reasoning regimes. We formulate and validate a switching linear dynamical system (SLDS) that reproduces these features, letting us simulate reasoning cheaply and anticipate critical transitions — misaligned states, adversarially-induced belief shifts, and other inference-time failures.

Presented at the 2nd Workshop on Reliable and Responsible Foundation Models (R2-FM) at ICML 2025.




Foundation Models for Multi-Omics

BARZILAY GROUP, MIT CSAIL — in progress

In the Barzilay group I build foundation models over multi-omics data — jointly modeling genomics, transcriptomics, and metabolomics — toward personalized identification of therapeutic targets. The goal is representations that integrate heterogeneous biological measurements well enough to surface candidate targets that single-modality models miss. Work ongoing; writeup to come.




Earlier work

Projects from high school, kept here for the record:

  1. general purpose geospatial compression (GPGC) — quad-tree raster decomposition that shrinks large-scale LiDAR elevation maps 20–200x with bounded error; open-source and used in certain NASA / VAIL systems.
  2. underfitting heuristic segmentation models — a deliberately bad hand-written heuristic for electrochemical-film defects that nonetheless trains a far superior neural segmentation model.
  3. a feature-generalizable technique for neural conditioning — pairing calcium-imaging readouts with an autoencoder to deliver targeted neurofeedback in rat motor cortex (MIT Jasanoff Lab).