Sun & Wu 2025 — Systematizing LLM Persona Design
Full title: Systematizing LLM Persona Design: A Four-Quadrant Technical Taxonomy for AI Companion Applications
Citation: Sun, E. & Wu, Z. (2025). Systematizing LLM Persona Design: A Four-Quadrant Technical Taxonomy for AI Companion Applications. Preprint under review, LLM Persona Workshop, NeurIPS 2025. arXiv:2511.02979v1
Core Contribution
Proposes a systematic framework for understanding AI persona applications along two axes:
- Interaction Intent: Emotional Companionship vs. Functional Augmentation
- Deployment Modality: Virtual vs. Embodied
This creates four quadrants, each with distinct technical requirements, design patterns, and ethical risks.
Key Arguments
- “Persona” is a fragmented concept, what works for virtual romantic companions differs fundamentally from enterprise assistants or physical robots
- Each quadrant has a core technical challenge: emotional consistency (Q1), reliability/safety (Q2), symbol grounding (Q3/Q4)
- Market bifurcation is occurring along these axes (e.g., mental health AI splitting into “wellness” vs. “clinical” tiers)
Extracted Content
Atoms:
- 05-atom—persona-drift
- 05-atom—empathy-paradox
- 05-atom—symbol-grounding-problem
- 01-atom—form-persona-dilemma
Molecules:
Notable Details
- Introduces a four-layer technical analysis framework: Model, Architecture, Generation, Safety & Ethics
- Documents the “writer-to-cultivator” shift in game design (from scripting to character seed creation)
- Identifies on-device SLMs as critical for real-time gaming applications due to latency constraints
- Argues that privacy-by-design is non-negotiable for embodied AI
Evaluation
Strengths: Comprehensive synthesis across disparate domains; useful organizing framework; practical implications for practitioners
Limitations: Preprint status; taxonomy may oversimplify edge cases; limited empirical validation of the framework itself
Related: 06-molecule—seci-framework, 07-molecule—ui-as-ultimate-guardrail