PhD in Computer Science at CSAIL & LFE · NSF Graduate Fellow
Zoe De Simone is a PhD student at MIT Computer Science & Artificial Intelligence Lab (CSAIL), and a member of the Laboratory for Financial Engineering, advised by Professor Andrew Lo, and an NSF Graduate Research Fellow. Her research focuses on human-AI collaboration, decision-making, and multi-agent systems, spanning generative AI alignment and the systemic risks of deploying AI at scale in financial markets.
She previously completed dual master's degrees in Computer Science and Building Technology at MIT, where she built decision-support tools for equitable city-scale building decarbonization, and earned her B.Arch from Cornell, where she co-developed Eddy3D and worked on surrogate ML models for airflow simulation.
She has worked in industry at Adobe Research, Point72, and SmartFab on applied AI systems spanning creative tools, finance, and manufacturing.
arXiv (2026)
Agent-based market simulations show that as LLM traders become more capable, their behavior grows more correlated. Shared reasoning lowers market risk when the agents are right, but amplifies it under common misinformation — a "capability paradox" where stronger components can weaken the overall system.
arXiv (2023)
Interactive web system for analyzing and editing bias in text-to-image diffusion models.
Energy Policy (2025) · MIT MS Thesis (2024)
Human-in-the-loop stochastic optimization framework for equitable retrofit planning.
Building Simulation Conference (2021)
Development of the Indoor Eddy3D simulation tool for floor plan layouts and ventilation impact on indoor air quality and pathogen spread. 20k+ downloads.
Forthcoming · Cornell Environmental Systems Lab
Physics-informed ML surrogate model reducing CFD simulation time from 3+ hours to ~1 second. Enables real-time, interactive airflow feedback from input encodings for early-stage design exploration.
arXiv (2026)
Creo is a multi-stage text-to-image system that supports ideation from sketch-like abstractions to high-fidelity outputs with stage-level edits and locking. In comparative studies, people reported stronger ownership and produced more diverse results than with one-shot generation.
arXiv (2026) · ICLR HCAIR Workshop
Human cognition moves from abstract to concrete, but instruction-tuned AI often short-circuits that process by jumping straight to a finished output, taking away user agency when goals are still forming. The paper names these failures "Fantasia interactions" and argues alignment research should train and evaluate models to allocate cognitive responsibility well within an interaction.
CVPR Workshop on Computer Vision in the Built Environment (2024)
Semantic segmentation pipeline for façade glazing ratio prediction from street-view imagery.
Journal of Building Engineering (2023)
Computational tool to assess building longevity through flexibility; practitioner survey (n=237).
Interned at Adobe Research
Worked on version control for AI/creative work.
Presented our paper, "Alignment has a Fantasia Problem," at the ICLR 2026 Workshop
From Human Cognition to AI Reasoning: Models, Methods, and Applications in Rio.
Workshop site →Presented a poster on Creo at the New England Symposium of Computer Graphics
Interned at Point72 as a quant in a macro pod for 6 months
Worked on generating trading dashboards and signals.
Started my PhD in Computer Science at MIT
Window-to-Wall Ratio Detection using SegFormer
Presented at the CVPR Workshop on Computer Vision in the Built Environment.
Paper →Graduated from MIT with a dual MS in EECS and Building Technology
Received the NSF GRFP for my work in CFD surrogate modeling for indoor spaces
Graduated with a BArch from Cornell with awards for best thesis