Sophie Broderick
About Me
Princeton CS 2027
Diversity of Generation in LLMs, Pluralistic Alignment, RL, Augmented LLMs
Research @
Reinforcement Learning Lab
Let’s talk about research or ideas for making adaptive intelligent systems that align with humanity. I also create art and write humor focused on the tech world, AI, and the tension between humans and machines.
Selected Research
-
NeurIPS 2026 LP4FM, WiML & Queer in AI poster | Escaping the Obvious: Reinforcement Learning for Conceptually Diverse GenerationSophie Broderick, Qiawen Liu, Benjamin EysenbachLarge Language Models often converge to a narrow set of high-probability ideas. This research investigates methods to mitigate this tendency, evaluating how RL changes the semantic geometry of model outputs for conceptually diverse generation.
Mahsa Bastankhah,Sophie Broderick,Benjamin EysenbachEmpowerment—an objective that measures how much influence an agent has over its environment—provides a principled way to learn concise representations that capture only the aspects of the environment that matter for control while ignoring irrelevant distractors.
This project evaluates the robustness of Maximum Entropy Reinforcement Learning (MaxEnt RL) via Energy-Based Normalizing Flows (EBFlow). While RL algorithms are typically tested in clean, idealized simulations, this work introduces realistic disturbances—including sensor corruption, action noise, and observation noise—using tools from the Robust Gym benchmark to assess generalization.
Published Humor Writing
Art
Interactive Art
Intimate Machines
Portraits
