Dr Sarah Mercer

Hi! I'm Sarah, a Software Engineer (formerly Principal Researcher at The Alan Turing Institute) working at the intersection of generative AI and agent-based systems.


  drsezzer |   Sarah Mercer

     The propensity of LLMs to portray human-like behaviour fascinates me. Since the publication of the Willowbrook report, I have continued to explore the capacity of generative agents to mimic human behaviour… exploring their ability to maintain believable and consistent personas, their capacity to make human-like mistakes, and their (in)ability to get angry!


DALLE-3 generated image of Willowbrook

     Inspired by the Stanford Smallville paper, a simulation comprising 12 characters and 10 locations - including a library, cafe, farm shop, village green and various residences - was developed to further explore the capacity of generative agents to portray human-like behaviours.

     Unlike Smallville, the Willowbrook simulation does not maintain a shared representation of the agents’ environment, meaning that the agents’ reality is purely LLM generated (with the exception of the initial character and location descriptions). This reality is held within each agent’s memory of what they observed, did and heard. As such, any error in the way the LLM is directing the agents is magnified as the simulation progresses and the agents’ memories of such inaccuracies are retained, or even acted upon. This provides a novel way to evaluate how different LLMs influence the lives of the Willowbrook residents.

     A common question is how generative agents are implemented and what frameworks are used. I don’t use a framework, as I wanted to avoid introducing an additional layer of abstraction between the system and the underlying model. The key to designing a good persona-agent lies in its initial biography and its memory retention mechanism.

Research Publications

Generative Agents / Simulated Societies:

Cyber Security / Protective Security:

Prior to working at the Turing, I was a researcher in Cyber Security. The interest garnered by LLMs at the beginning of 2023 obviously had an impact on the cyber security community. The paper below, was my attempt to bring some evidenced thinking to the fairly polarised (at the time) debate, given my familiarity of developing LLM based applications and intuition for their strengths and weaknesses. Note: Technical readers may prefer the unedited version of the paper, as linked below.

Other Topics:

CETaS papers:

Alongside my own research looking at the human like capacity of generative agents, I also provided technical expertise to the CETaS team, specifically Generative AI.

All rights are reserved for the contents on this site (drsezzer.github.io), same for uploaded PDFs unless otherwise stated within the document itself.
(c) 2026 Sarah Mercer.

Web site generated using GitHub Pages with Jekyll Theme Midnight and using the jemoji plugin (emoji cheat sheet).