Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations
Rotem Dror, Zohar Elyoseph, Yuval Haber, Elad Refoua, Oshrat Ayalon, Adir Solomon
University of Haifa · Bar-Ilan University
Published September 6, 2026 · on the feed September 22, 2026 · ▲ 22 · arXiv 2609.22255
The abstract
Existing approaches to persona simulation with Large Language Models (LLMs) mostly rely on shallow character descriptions that fail to sustain coherent character behavior across extended interactions. We introduce Deep Persona, a psychologically grounded, three-layered architecture that organizes personas into hierarchical levels of observable expression, latent beliefs, and core motivational drives, for constructing highly convincing role-playing agents. Governed by the principles of scripted determinism and bounded agency, the architecture restricts the model to a reactive engine guided by a structured internal script. We further propose a reference-free evaluation framework that benchmarks dialogue naturalness against empirical human distributions using established psychological clinical instruments and adversarial stress-tests. Empirical evaluation reveals that while LLMs achieve high pragmatic fluency, they exhibit systematic limitations in emotional expression and joint attention. In addition, we present a case study of two Deep Personas and evaluate them using the proposed framework, demonstrating that structured personas can produce interactions that more closely align with human conversational behavior.
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
Deep Persona. A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations. By Rotem Dror and five others, from the University of Haifa and Bar-Ilan University. Published on the sixth of September. No code repository is linked to it.
Most persona simulation with large language models starts from a character description, and those descriptions are shallow. Over a long interaction, a thin sketch is not enough to keep a character behaving coherently. Somewhere in an extended exchange, the character stops behaving like one person.
Deep Persona gives the character depth in a psychologically grounded way. It is a three-layered architecture, organized as a hierarchy. At the surface is observable expression. Beneath that are latent beliefs. At the base are core motivational drives. The aim is role-playing agents that are highly convincing because what they say is anchored to what they believe and want.
Two principles govern the architecture: scripted determinism, and bounded agency. Under them, the model is restricted to acting as a reactive engine, guided by a structured internal script. The language model supplies the voice. The script supplies who the character is.
Alongside the architecture comes an evaluation framework that needs no reference. It benchmarks how natural the dialogue is against empirical human distributions, using established clinical instruments from psychology, and it applies adversarial stress-tests.
The empirical evaluation finds a split. Large language models achieve high pragmatic fluency, but they show systematic limitations in two areas: emotional expression, and joint attention. The paper also presents a case study of two Deep Personas, evaluated with the proposed framework. It reports that structured personas can produce interactions that align more closely with human conversational behavior.
The contribution comes in two parts that serve each other. One is a way to build a persona with layers beneath the surface, kept on script. The other is a way to measure the result against how people actually converse. The measurement is also what exposes where the models still fall short, in emotion and in shared attention, even when their fluency is high.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A character who forgets who they are halfway through a session. That is the wall here.
You are building a language-model character to help train therapists: a patient who must stay the same person through a long, difficult session.
The road ahead
- 1 · The problem ~6 min A character written as a short list of traits sounds right at first, then drifts and contradicts itself as a long conversation goes on.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Build the character in three layers, what it shows, what it believes and what drives it, and let the model only voice that script.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Write the layered persona, run the model as a reactive engine on that script, then test it against human dialogue and adversarial attacks.
- 6 · The experiment ~6 min You’ll guess: Measured against human dialogue, where did large language models show systematic limitations?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.