Stranger, Fan, or Peer? A Systematic Study on the Role of Interlocutor in Persona-Based Dialogue Generation

The specific study titled "Stranger, Fan, or Peer? A Systematic Study on the Role of Interlocutor in Persona-Based Dialogue Generation" that factorizes biography visibility across training, inference, and evaluation stages is not present in the provided search context.

The available search results focus on the paper "When Harry Meets Superman: The Role of The Interlocutor in Persona-Based Dialogue Generation" (published May 2025), which investigates interlocutor impact through a different methodology:

• Evaluation Framework: It uses a speaker identification task where evaluators (human and LLM-as-a-judge) attempt to identify a target speaker's biography from generated dialogue.
• Information Masking: It systematically masks or discloses the interlocutor’s biographical information and dialogue turns during evaluation to measure their influence on the target speaker's persona consistency.
• Key Findings:
• Interlocutor Importance: Access to the interlocutor’s biography significantly improves the recognition of the target speaker, indicating models adapt responses to the interlocutor's profile.
• Familiarity: Models generalize well across topics but struggle with unfamiliar interlocutors compared to familiar pairings.
• Zero-Shot vs. Fine-Tuned: Zero-shot models often copy biographical details directly (leading to superficial dialogues), while fine-tuned models capture deeper speaker traits more effectively.

The paper investigates persona-based dialogue generation as a multi-stage social-information problem rather than a simple conditioning task. It asks how a model’s use of a character’s biography changes depending on two factors: the interlocutor role—stranger, fan, or peer—and the visibility of biography information at different points in the pipeline, including training, inference, and evaluation. This framing is important because prior work often treats persona as a static prompt or profile, or evaluates persona consistency in isolation, which can obscure whether a model is truly using persona information, merely retrieving it, or benefiting from evaluation-time leakage.

A key contribution is the paper’s factorized experimental design, which separates variables that earlier studies typically confound. By varying when biography information is available and to whom the model is speaking, the work exposes mechanisms hidden by single-factor treatments. For example, apparent persona consistency may depend heavily on whether the biography is present in the inference context, whether the model was trained on persona-rich data, or whether the evaluation itself rewards persona-like surface features. The interlocutor role further modulates this behavior, since a fan, peer, and stranger may elicit different levels of self-disclosure, persona assertiveness, and social accommodation.

This matters because persona-based dialogue is increasingly used in character chatbots, social simulation, and interactive assistants, where reliable evaluation and socially appropriate behavior are essential. The study suggests that future systems should explicitly model social relationship and information availability, not just persona content, and that benchmarks should report the visibility conditions under which persona performance is measured. More broadly, it reframes persona generation from profile adherence to context-sensitive social reasoning, offering a more rigorous foundation for both model design and evaluation.