Researchers increasingly use large language models across the research workflow, from ideation to qualitative analysis and participant simulation. A new preprint on arXiv (2610.27333) focuses on one slice of that trend: using LLMs to rehearse qualitative interviews. The authors argue that prior work has largely evaluated simulated participants on outcome measures, without much attention to what is happening inside the simulation.
To address this, they propose 'inspectable' simulated participants—designs that let researchers see how and why the simulated person responds as they do. The goal is not to replace human interviewees but to give interviewers a transparent, low-stakes way to practice before going into the field.
Because this is a single preprint, the claims are not yet peer-reviewed, and no comparison with other approaches is available from this source alone. Still, the emphasis on inspectability points to a broader concern: as LLMs take on more research roles, understanding their internal behavior becomes as important as the outputs they produce.