Dynamic In-Group Persona Generation for Enhancing Human-AI Rapport

Technical Summary: Dynamic In-Group Persona Generation for Enhancing Human–AI Rapport

Problem Statement
While Large Language Model (LLM)-based chatbots are increasingly deployed in interpersonal domains such as counseling and peer support, establishing genuine human–AI rapport remains a significant challenge. Rapport, defined as mutual attention, positivity, and coordination, is essential for user trust and sustained engagement. Current systems often fail to demonstrate the contextual understanding of social cues required to form rapport comparable to human counselors. Existing approaches to building rapport rely on shallow similarity (e.g., tone matching) or generic self-disclosure (e.g., "I've felt that too"), which often lack the depth of shared lived experiences or social identity necessary for deep relational connection. Furthermore, prior studies have not sufficiently explored how contextual alignment with a user's specific concerns can be operationalized to enhance agent effectiveness.

Methodology
The authors propose the In-group Persona Agent (IPA), a framework that dynamically generates synthetic personas grounded in a user's expressed concerns and context. The methodology operates through a multi-stage pipeline:

1. Pre-chat Dialogue Elicitation: A "collector" agent engages the user to identify their primary concern (e.g., career anxiety) and brief personal context. The system performs a sufficiency check every two turns; if information is insufficient, it asks clarifying questions until a sufficient profile is established.
2. Dynamic Persona Generation: Based on the collected information, the system generates five distinct synthetic in-group personas.
   • Shared Concern: All generated personas share the user's primary concern.
   • Divergent Background: The personas differ in background details (e.g., age, profession, specific narrative) to provide diverse perspectives while maintaining relevance.
   • Structure: Each persona consists of a Background (history, education, interests) and a Narrative (a story of facing a similar challenge and resolving it).
   • Selection: The user selects one of the five generated profiles before the main conversation begins.
3. Conversational Inference: The main dialogue agent is conditioned with the selected persona profile. This enables the agent to produce responses that include strategic, context-aligned self-disclosure and empathetic narratives, rather than generic advice.

The study employs a randomized control trial (RCT) with 170 participants (after exclusions) across three conditions:

• No Persona (NoP): A standard LLM agent with no persona conditioning.
• No Persona + Self-Disclosure (NoPs): An agent without a persona that uses brief, generic self-disclosure statements not tailored to the user's specific context.
• In-group Persona Agent (IPA): The proposed system with dynamic, concern-aligned persona generation.

Key Contributions

• In-group Persona Agent (IPA): A novel method that integrates persona generation and conversational inference via a multi-stage prompt pipeline, constructing personas dynamically from elicited user dialogue rather than static backstories.
• Systematic Evaluation: A human-subject study comparing IPA against conventional agents and generic self-disclosure baselines, utilizing post-task questionnaires and turn-level behavioral analysis.
• Validation Framework: The introduction of an LLM-based evaluation rubric to assess persona quality (In-group Fitness and Concern Resolution Quality) and a sufficiency check mechanism to verify the adequacy of pre-chat information for persona generation.

Results
The study utilized rank-based non-parametric statistical tests (Jonckheere–Terpstra and Mann–Whitney U) to analyze outcomes.

• Rapport: The IPA condition significantly outperformed both baselines across all rapport metrics. The improvement was most pronounced in items measuring perceived personal relevance ("relevant to me"), bonding ("bond with the agent"), and personal interest. The generic self-disclosure baseline (NoPs) showed only modest benefits over the No Persona baseline, suggesting that alignment is more critical than disclosure alone.
• User Experience (UX): IPA yielded the highest scores overall, with the most significant gains observed in engagement and perceived relevance. The NoPs condition showed mixed results compared to NoP, indicating that generic self-disclosure does not reliably improve the overall user experience.
• Behavioral Analysis: Turn-level analysis revealed that IPA uniquely strengthened self-disclosure reciprocity. Users were significantly more likely to engage in deep self-disclosure following deep agent self-disclosure in the IPA condition compared to other conditions. Additionally, agent empathy following user self-disclosure was highest in the IPA condition.
• Correlation: Higher persona quality scores (specifically "Shared Skills") correlated significantly with the user's intention to continue the conversation, though persona quality did not comprehensively explain all rapport dimensions.

Significance and Claims
The paper claims that embedding concern-aligned in-group personas into conversational AI is a practical and effective method for strengthening rapport and engagement in supportive interpersonal settings. The authors posit that the key mechanism is not merely self-disclosure, but the contextual alignment of the agent's identity with the user's specific emotional and goal-oriented states.

The study demonstrates that lightweight, prompt-based persona design can bridge the gap between generic AI responses and the relational depth required for peer support. However, the authors maintain a modest scope, noting that their findings are specific to lower-risk domains like career counseling. They explicitly caution against generalizing these results to high-stakes clinical settings (e.g., suicide prevention) without further safeguards, acknowledging that synthetic empathy in sensitive domains could lead to unintended consequences such as undue trust or the discouragement of professional help-seeking. The work highlights in-group alignment as a critical factor in human–AI relational quality, offering a pathway to more empathetic and engaging AI interactions.