Researchers have introduced Deep Persona, a psychologically grounded architecture and evaluation framework designed to make role-playing agents built on large language models more coherent over extended interactions. The work addresses a common weakness in existing persona simulations: shallow character descriptions that struggle to sustain consistent behavior across long conversations.
The architecture organizes each persona into three hierarchical layers: observable expression, latent beliefs, and core motivational drives. It is governed by two principles — scripted determinism and bounded agency — which restrict the model to acting as a reactive engine guided by a structured internal script rather than improvising freely.
Alongside the architecture, the authors propose a reference-free evaluation framework that benchmarks dialogue naturalness against empirical human distributions, drawing on established psychological clinical instruments and adversarial stress-tests.
Empirical evaluation found that while LLMs reach high pragmatic fluency, they show systematic limitations in emotional expression and joint attention. A case study of two Deep Personas evaluated under the proposed framework indicated that structured personas can produce interactions that more closely align with human conversational behavior.