EVALUATING PERSONALITY-DRIVEN LARGE LANGUAGE MODELS AS SIMULATED USERS IN INTERACTIVE APPLICATIONS
=================================================================================================

Can a language model, given a personality profile and interface context, behave like a realistically different user?

I studied LLM-driven agents interacting with ten websites through Playwright. The agents received personality-conditioned instructions based on the Big Five model, then navigated interfaces through actions such as clicking, entering text, and scrolling.

I logged their actions, page visits, exploration behavior, and failures to test how consistently the simulated personalities led to distinct behavior. The results exposed limitations in treating personality-prompted LLMs as substitutes for real people. This was not a human-subject validation study.

Tools: Python, Playwright, LLM APIs, pandas

Publication:
  Isaac C. Robinson (2026)
  International Journal of High School Research, Volume 8, Issue 11, pages 75–80
  [Read the paper](https://doi.org/10.36838/IJHSR811.75)

Code:
  [Research code](https://github.com/crbns/llm-synthetic-user-sessions)