Stranger, Fan, or Peer? A Systematic Study on the Role of Interlocutor in Persona-Based Dialogue Generation
This study systematically analyzes how interlocutor biography visibility across training, inference, and evaluation affects persona-based dialogue generation, revealing leakage mechanisms.
Key Findings
Methodology
Using the PRODIGy dataset, the study introduces a multi-stage visibility framework, controlling target and interlocutor biography access during training, inference, and evaluation. Fine-tuning LLaMA 3.1 8B with LoRA adapters under three configurations (strangers, fans, peers), the research assesses how different visibility settings influence dialogue content, especially biographical copying. An LLM judge performs author identification to quantify information leakage. The analysis employs metrics like rare-word overlap, accuracy, and diversity, revealing the dominant influence of training-stage visibility on model behavior and leakage pathways.
Key Results
- Training with interlocutor biography visibility reduces direct copying (rare-word overlap drops from 41% to below 5%) while maintaining high identification accuracy (up to 97.1%). Inference-time information improves recognition but causes less consistent leakage, especially under asymmetric disclosure. When only the interlocutor sees the target biography, target content more often leaks into interlocutor turns, making identification easier. These results demonstrate that multi-stage configuration critically influences the balance between persona fidelity and privacy leakage.
- Models trained with interlocutor information better express consistent personas and reduce surface copying. Introducing information at inference enhances recognition but can increase leakage, especially if training did not expose the model to such information. The findings highlight that training-stage exposure is key to controlling copying, while inference-stage information can be used to improve performance without necessarily increasing leakage if configurations are aligned.
- Overall, the experiments establish that separating training, inference, and evaluation stages allows precise control over information flow, enabling the design of dialogue systems that balance authenticity and privacy. The results provide a foundation for developing safer, more controllable persona-based models, with implications for privacy-preserving AI and personalized interaction.
Significance
This research advances understanding of multi-stage information dynamics in persona dialogue systems, addressing critical issues of privacy leakage and persona consistency. By dissecting how different stages influence model behavior, it offers a blueprint for designing safer AI systems that can maintain personalized interactions without compromising sensitive data. The multi-stage framework bridges a gap in existing literature, which often considers only static or single-stage settings, thus opening new avenues for privacy-aware dialogue generation. Its insights are valuable for deploying AI in sensitive domains like healthcare, counseling, and customer service, where privacy and authenticity are paramount. Moreover, the methodology sets a precedent for future research on multi-stage information control, fostering safer AI development.
Technical Contribution
The paper introduces a novel multi-stage (training, inference, evaluation) framework for analyzing biography visibility effects in persona dialogue models. It combines controlled data generation with fine-tuned LLaMA models and a dedicated LLM judge for author attribution, enabling precise quantification of information leakage pathways. The approach disentangles the influence of each stage, revealing that training exposure predominantly governs copying behavior, while inference information can enhance recognition without necessarily increasing leakage if aligned with training conditions. The methodology employs multiple metrics, including rare-word overlap and diversity measures, to comprehensively evaluate model behavior, providing a new analytical paradigm for privacy and persona fidelity in dialogue systems.
Novelty
This work is the first to systematically separate and analyze the effects of biography visibility across training, inference, and evaluation stages in persona-based dialogue generation. Unlike prior studies that only assess information leakage at the evaluation stage, this research highlights the dominant role of training exposure in shaping model behavior and leakage pathways. The multi-stage framework and the use of an LLM judge for author identification represent innovative steps, offering a deeper understanding of how information propagates and leaks in dialogue models. This approach provides new insights into designing privacy-aware, persona-consistent AI systems.
Limitations
- 研究主要基于电影角色数据集，实际应用中多模态、多领域场景的表现尚未验证，模型在复杂环境下的鲁棒性有限。
- 多阶段配置增加系统复杂性，可能影响部署效率和实时性。
- 在极端信息泄露场景下，模型仍可能出现人格偏差，未来需结合差分隐私等技术优化。
Future Work
未来将结合多模态信息（如视觉、声音）扩展多阶段分析，探索差分隐私和联邦学习技术，提升模型的隐私保护能力。同时，研究将面向更复杂的实际应用场景，验证多阶段配置的实用性和安全性，为人格对话系统的商业化提供理论支持。
AI Executive Summary
This study tackles the critical challenge of managing information leakage in persona-based dialogue systems by proposing a comprehensive multi-stage framework. Traditional approaches often focus solely on evaluation metrics, overlooking how information flows during training and inference stages. Here, the authors introduce a novel methodology that explicitly controls the visibility of speaker biographies across these three phases, enabling a systematic analysis of how such information influences dialogue content and privacy. Using the PRODIGy dataset, they design three configurations—strangers, fans, and peers—to simulate different real-world scenarios, fine-tuning LLaMA 3.1 8B models with LoRA adapters for each setting.
The core insight is that training-stage exposure to interlocutor information predominantly determines the extent of biographical copying and persona expression. When models are trained with access to interlocutor biographies, they produce more consistent and less surface-level content, reducing direct copying from biographies. Conversely, in settings where such information is absent during training, models tend to copy biographical details verbatim, risking privacy leaks.
In the inference phase, introducing biography information can improve the model’s ability to recognize and express personas, but it also increases the risk of leakage, especially under asymmetric disclosure where only one participant knows the other's biography. The experiments demonstrate that aligning training and inference configurations minimizes leakage while maintaining high recognition accuracy, highlighting the importance of multi-stage design.
This research offers a significant step forward in understanding and controlling information flow in dialogue systems. Its implications extend to privacy-preserving AI, personalized virtual assistants, and sensitive data handling, providing a blueprint for safer, more reliable AI interactions. Future work will explore integrating advanced privacy techniques like differential privacy and expanding to multi-modal data, aiming to develop robust, privacy-aware conversational agents that can operate safely across diverse real-world scenarios.
Deep Dive
Abstract
Persona-based dialogue systems are usually conditioned on speaker biography, but dialogues involve at least two participants, and who has access to whose biography can vary across training, inference, and evaluation. Prior work often neglected these aspects, obscuring mechanisms that only appear when biography visibility is toggled separately across training, inference, and evaluation, a three-stage factorisation that prior work has largely treated as a single factor. We study this factorisation on a dataset of dialogues paired with speaker's biographies, varying whether the target and interlocutor speakers see each other's biographies during training and inference, and using an LLM as a judge to perform author identification. We find that (i) training-time visibility, more than inference-time visibility, determines whether models express persona traits through dialogue or fall back on copying biographical text (a known problem/phenomenon in persona-based generation); (ii) models trained with interlocutor-biography visibility copy less target-biographical text than models trained without it, while changing visibility only at inference time has a less consistent effect; and (iii) under asymmetric disclosure, where only the interlocutor sees the target biography, target content leaks into interlocutor turns more often, and dialogues containing such traces are easier for the judge to identify, especially when interlocutor turns are visible. These results suggest that biography leakage into generated turns is an artefact of how interlocutor visibility is configured across training and inference, and separating the three stages is necessary.