Nathaniel Morgan

Cambridge, MA

Nathaniel Morgan is a Computer Science and Artificial Intelligence student at MIT (Class of 2027) focused on building intelligent systems that bridge research and real-world deployment. He is currently an undergraduate researcher at MIT CSAIL in the Spoken Language Systems Group, where he explores retrieval-augmented generation and latent model editing techniques to improve long-term memory in large language models.

In industry, Nathaniel has worked at startups including SubconsciousAI, BitEnergy.AI, and Salieri.AI, contributing to projects involving LLM inference engines, DevOps infrastructure, and AI research. He also leads ARK, an open-source platform for modular LLM applications and scalable inference. His work spans serverless deployment, container orchestration, and toolchain integration for applied machine learning systems.

Nathaniel is the founder of Uplift.co, a 501(c)(3) nonprofit dedicated to youth mental wellness. He has organized four in-person events and launched over eight digital wellness campaigns. He also serves on the Board of Directors of the Harvard-MIT COOP, where he helps develop connect the COOP to its comunity as Co-Chair of the Brand Experience comittee.

Outside of his technical and leadership work, Nathaniel enjoys longboarding, soccer, violin, piano, photography, and videography as creative outlets that inform his approach to design and problem-solving.

You can reach him at: nmorgan [at] mit [dot] edu

selected publications

Thinking Deeper with Recursive Spawning (ThReaD)

NAACL, 2025

arXiv:2405.17402 [cs.CL]

Abs

Large language models (LLMs) have demonstrated impressive capabilities across diverse settings but still struggle as the length and complexity of the context increases. To address this challenge, we propose Thinking Recursively and Dynamically (ThReaD), a framework that adapts model generation by dynamically spawning new threads based on context. This method decomposes tasks into simpler sub-problems, enabling recursive problem-solving for complex tasks. ThReaD outperforms existing methods on benchmarks including ALFWorld, TextCraft, WebShop, DataCommons QA, and MIMIC-III ICU QA.