Why do static persona descriptions produce repetitive dialogue?

Standard persona-based dialogue datasets (PersonaChat, Synthetic Persona Chat, Blended Skill Talk) define personas through static, predefined descriptions — typically 3-5 attribute sentences like "I have two dogs" or "I work as a nurse." The Beyond Discrete Personas study documents three failure modes:

1. Repetitiveness: conversations loop back to the same persona attributes
2. Shallowness: dialogue stays at the surface level of stated facts
3. Contradiction: the model generates responses that conflict with its own persona description

The proposed alternative: instead of static attribute lists, use long-form journal entries — authentic, unfiltered self-expression from platforms like Reddit — to capture personality dynamically. The approach clusters journal entries per author, filters for representativeness, and maps them to Big Five personality traits.

This produces ~400,000 dialogues where personality emerges from the way people describe their own experiences, thoughts, and emotions — not from a list of facts about them. The distinction matters because personality is not a set of attributes but a pattern of expression. Static descriptions capture what someone is; journal entries capture how they think and feel.

The connection to Can AI agents learn people better from interviews than surveys? is direct: both findings converge on richness of self-expression as the key variable. Interviews and journal entries share the property of being extended, authentic, unstructured personal narrative — the opposite of attribute lists.

The practical design principle: persona systems should be seeded with extended naturalistic text from the target individual, not condensed attribute descriptions. The more a persona description resembles a database record, the worse the simulation.

Tree-structured persona maintenance for multi-turn stability (from Arxiv/Agents Multi): The CGMI framework identifies a specific failure mode that static personas exacerbate: LLMs tend to forget original character settings in multi-turn dialogues and make decisions inconsistent with the character's design. Additionally, context window limitations make comprehensive fine-detailed role-setting challenging. The solution is a tree-structured persona model for character assignment, detection, and maintenance — organizing personality attributes hierarchically so that core traits anchor subordinate behaviors. Combined with an ACT-inspired cognitive architecture (Adaptive Control of Thought) that uses Chain of Thought and Chain of Action to extract declarative and procedural memories from working memory, this ensures "deeper and more specialized insights" during reflection and planning. The tree structure enables systematic detection of persona drift — when generated behavior deviates from a branch of the persona tree, the system can identify *which aspect has drifted and correct specifically, rather than re-prompting the entire persona description.