Overview
This is a field survey: it synthesizes existing work and points to methods and gaps, but does not provide new experimental benchmarks or single definitive solutions.
Citations3
Evidence Strength0.60
Confidence0.80
Risk Signals12
Trust Signals
Findings with numeric evidence: 1/6
Findings with evidence refs: 6/6
Results with explicit delta: 0/0
Reproducibility
Status: Partial assets available
Open source: Partial
At A Glance
Cost impact: 60%
Production readiness: 40%
Novelty: 50%
Why It Matters For Business
Personas let LLMs act like domain experts or tune results to users; use prompt personas to quickly prototype role-based workflows and invest in privacy-safe personalization for customer retention.
Who Should Care
Summary TLDR
This survey organizes work that uses "persona" with large language models into two clear streams: (1) role-playing, where the model is given a persona or role to act as; and (2) personalization, where the model encodes and uses a user's persona to tailor outputs. It catalogs environments (software development, games, medical, evaluation), methods (prompt personas, multi-agent frameworks, retrieval for long histories), evaluation tools (Big Five, MBTI, MPI, LLM-as-evaluator), and open problems (long-context memory, dataset gaps, bias, safety, privacy). The paper provides a practical map, representative systems, and future directions.
Problem Statement
Research on using personas with LLMs is growing but fragmented. Practitioners lack a unifying taxonomy, a clear mapping from use cases to methods, and a concise view of evaluation practices and safety/privacy gaps.
Main Contribution
A two-part taxonomy: LLM Role-Playing (model has persona) vs LLM Personalization (model adapts to user persona).
A review of environments and methods: prompts, multi-agent agents, retrieval-memory, and fine-tuning approaches.
Key Findings
The field splits into two distinct goals: role-playing and personalization.
Role-playing is often effective with prompt-based, training-free methods.
What To Try In 7 Days
Prototype a persona prompt for a concrete role (support agent, medical reviewer) and test outputs.
Build a simple 2–3 agent pipeline (planner + worker + reviewer) for a multi-step task.
Run a Big Five quick test on role-play outputs and compare to expected traits with a small human panel.
Agent Features
Memory
Planning
Tool Use
Frameworks
Is Agentic
Yes
Architectures
Collaboration
Reproducibility
Risks & Boundaries
Limitations
Heterogeneous metrics across subfields make direct comparisons hard.
Many tasks lack standardized datasets or format-specific benchmarks.
When Not To Use
When legal privacy constraints forbid storing user persona data in prompts or memory.
When strict safety or non-toxicity guarantees are required without additional safeguards.
Failure Modes
Bias amplification when assigning demographic personas.
Increased toxicity or harmful outputs under certain persona prompts.