PaperlayerView exact revision on arXiv

Source-linked AI summary

PersonaBench: Evaluating AI Models on Understanding Personal Information through Accessing (Synthetic) Private User Data

Juntao Tan, Liangwei Yang, Zuxin Liu, Zhiwei Liu, Rithesh Murthy, Tulika Manoj Awalgaonkar, Jianguo Zhang, Weiran Yao, Ming Zhu, Shirley Kokane, Silvio Savarese, Huan Wang, Caiming Xiong, Shelby Heinecke

arXiv:2502.20616v2cs.AI

TL;DR

Personalized assistants need to understand users’ personal attributes. PersonaBench uses synthetic profiles and private documents to benchmark that capability, and evaluations find existing RAG pipelines insufficient for personal questions.

  • Problem

    Personalized AI responses require models to understand users’ biographical details, preferences, and social connections.

    Source: §1 · p. 1

  • Method

    PersonaBench builds diverse synthetic user profiles and private documents, then benchmarks models’ understanding of personal information from that data.

    Source: §1 · pp. 1–2, §6 · p. 8

  • Results

    Evaluations indicate that existing RAG pipelines are insufficient for effectively answering personal questions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings point to a need for more sophisticated systems that extract and use personal information from noisy, fragmented, and irrelevant data.

    Source: §6 · p. 8

    Synthetic documents may contain occasional inconsistencies or unrealistic details, and perfect coherence cannot be guaranteed.

    Source: §7 · p. 9

Abstract

from arXiv · show

Personalization is critical in AI assistants, particularly in the context of private AI models that work with individual users. A key scenario in this domain involves enabling AI models to access and interpret a user's private data (e.g., conversation history, user-AI interactions, app usage) to understand personal details such as biographical information, preferences, and social connections. However, due to the sensitive nature of such data, there are no publicly available datasets that allow us to assess an AI model's ability to understand users through direct access to personal information. To address this gap, we introduce a synthetic data generation pipeline that creates diverse, realistic user profiles and private documents simulating human activities. Leveraging this synthetic data, we present PersonaBench, a benchmark designed to evaluate AI models' performance in understanding personal information derived from simulated private user data. We evaluate Retrieval-Augmented Generation (RAG) pipelines using questions directly related to a user's personal information, supported by the relevant private documents provided to the models. Our results reveal that current retrieval-augmented AI models struggle to answer private questions by extracting personal information from user documents, highlighting the need for improved methodologies to enhance personalization capabilities in AI.

1 Introduction

The paper addresses the lack of public resources for assessing personal-information extraction from private user data by introducing a synthetic-data pipeline and benchmark. Its evaluation finds that existing RAG pipelines are insufficient for answering personal questions from those documents.

1 Introduction

  • Privacy concerns leave few publicly available user documents paired with ground-truth personal information, making objective evaluation difficult.Personal information may also be noisy, fragmented, or change over time.Source: §1 · p. 1
  • The pipeline creates diverse synthetic profiles and private documents that reveal personal attributes for standardized evaluation.The profiles serve as ground truth for generating documents such as conversation histories, AI interactions, and purchase histories.Source: §1 · pp. 1–2
  • Generating realistic data is challenging because attributes must remain diverse, relationships natural, and documents authentic despite noise and changing information.These are identified as key complexities in creating the dataset.Source: §1 · p. 2
  • The evaluation keeps profiles hidden and requires models to extract personal information solely from accessible private documents.The pipeline integrates persona sampling, social graph creation, profile completion, and multiple private-document types.Source: §1 · p. 2
  • Existing RAG pipelines are insufficient for effectively answering personal questions using private user documents.The authors argue that more advanced methodologies and systems are needed to enhance personalization.Source: §1 · p. 2

2 Related Works

Related work covers persona-grounded generation and benchmarks for personalized generation. The paper distinguishes its focus on extracting multifaceted personal information from work centered on dialogue, persona sets, API use, or stylistic imitation.

2.1 LLMs for Persona Modeling and Grounded Generation

  • Persona-grounded generation has progressed from persona-based dialogue to prompting for profile-grounded conversations; this work extends it to multiple document types and longer time spans.Other LLM-based simulations have generated self-reports, questionnaire responses, and social interactions.Source: §2.1 · pp. 2–3

2.2 Evaluating LLMs on Personalized Generation

  • Tau-bench and AppWorld emphasize fulfilling requests through external APIs, while LaMP focuses on style or behavioral imitation rather than extracting multifaceted personal information.The paper positions its focus as accurate extraction and interpretation of complex personal attributes.Source: §2.2 · p. 3

3 Stage 1: User Profile Synthesis

PersonaBench constructs comprehensive synthetic profiles and social communities, then completes each profile in stages so that personal attributes remain grounded in persona and social context.

3 Stage 1: User Profile Synthesis

  • Profiles represent varied characters connected in communities and serve as the source for generating private documents and personalized questions and answers.The profiles include attributes such as occupations, eating habits, and favored activities.Source: §3 · p. 3

3.1 Profile Template Definition

  • Each profile organizes demographic, psychographic, and social information, while allowing user-specific subcategories to vary.The template builds on a hierarchical user-template approach introduced in Lee et al. (2022).Source: §3.1 · pp. 3–4

3.2 Persona Sampling and Social Graph Creation

  • Sampled persona descriptions and generated names are used to reduce duplication, while a social graph is built and checked for symmetric, consistent relationships.The graph starts with three sampled personas and is expanded with additional connected individuals.Source: §3.2 · p. 4

3.3 Profile Completion

  • Profile completion first generates attributes grounded in social relationships, then fills remaining attributes using each character’s persona and existing details.The two steps are designed to support social consistency and an internally consistent, well-rounded profile.Source: §3.3 · p. 4

4 Stage 2: Private Data Synthesis

The pipeline synthesizes three types of private documents and varies their content to model personal disclosures, irrelevant noise, real-world context, and changing preferences over time.

4 Stage 2: Private Data Synthesis

  • Private data synthesis produces conversations, user–AI interactions, and purchase histories to reflect characters’ behaviors and daily activities.The generation strategy also accounts for noisy data, real-world events, and possible changes to personal information.Source: §4 · p. 4, §4.2 · p. 5

4.1 Document Types

  • Conversation records draw on socially connected individuals, user–AI interactions capture direct chats, and purchase histories reflect persona preferences.Purchase records retain selected item features, and document sessions include timestamps and may link contextually across time.Source: §4.1 · pp. 4–5

4.2 Generation Strategy

  • The strategy combines attribute-revealing sessions with noise, real-world events, and occasional updates to personal information.Each document type contains only part of a person’s information, so understanding the individual requires combining documents.Source: §4.2 · p. 5, §4.1 · p. 5

4.2.1 Personal Data Generation

  • Personal-data sessions are prompted to subtly reveal a randomly selected attribute, and conversations may mention attributes of people in the user’s social network.The prompts discourage explicit, direct statements.Source: §4.2.1 · p. 5

4.2.2 Noise Data Generation

  • Noise sessions cover general topics without revealing participants’ personal traits or preferences, and the noise-to-personal-data ratio is controlled.The authors expect higher noise levels to reduce retrieval accuracy.Source: §4.2.2 · p. 5

4.2.3 Real-world News Integration

  • Conversation prompts have a 20% chance of including a public news event aligned with the time window of the user’s private data.The articles are retrieved using an external tool.Source: §4.2.3 · p. 5

4.2.4 Information updating

  • Less than 1% of generated cases update a previously mentioned preference in a later conversation to model changing attributes.The authors describe this evolution as increasing dataset complexity.Source: §4.2.4 · p. 5

5 Experiments

Across the tested RAG systems, retrieval and end-to-end personal question answering remain challenging, and both worsen as document noise increases. The experiments compare model configurations, question types, and performance dimensions.

5 Experiments

5.1.2 Private Documents Statistics

  • Private documents are generated at noise ratios of 0.0, 0.3, 0.5, and 0.7, with Table 1 reporting sessions and utterances by document type.For each person, queried attributes are guaranteed to appear, while other attributes may be included or omitted.Source: §5.1.2 · p. 6, Table 1 · p. 6

5.1.3 Personal Q&A Statistics

  • Questions cover demographic, psychographic, and social attributes; demographic and psychographic attributes are downsampled, while all social attributes are queried.Social questions can be multi-hop, such as identifying a sister before finding her favorite movie.Source: §5.1.3 · p. 6

5.2 Model Selection and Implementation Details

  • The evaluation combines four GPT base models with three dense retrievers, yielding 12 distinct RAG models.The retrievers include all-MiniLM-L6-v2, all-mpnet-base-v2, and bge-m3.Source: §5.2 · pp. 6–7

5.3 Evaluation Metrics

  • Retrieval is measured with Recall and NDCG, while end-to-end answer correctness is measured with Recall and F1-score.End-to-end metrics compare predictions with reference answers that often contain multiple distinct terms.Source: §5.3 · p. 7

5.4 Results

  • At 50% noise, bge-m3’s overall Recall is 0.325; GPT-4o-mini achieves the best overall end-to-end results, while ground-truth context still yields Recall around 50%.bge-m3 outperforms all-mpnet-base-v2 and all-MiniLM-L6-v2 by 21.7% and 37.7% in Recall, respectively; the ground-truth-context result suggests some information is conveyed implicitly.Source: Table 3 · p. 7, Table 4 · p. 7, §5.4 · p. 8
  • As noise increases, retrieval performance consistently declines, along with overall RAG performance.The study attributes the end-to-end decline in part to reduced retrieval performance.Source: Figure 3 · p. 6, Figure 4 · p. 7, §5.4 · p. 8
  • Across six dimensions, GPT-4o-mini leads overall but is weaker on updating and noise sensitivity, while GPT-4o performs better on those aspects; bge-m3 is well-rounded.The dimensions include basic and social information, noise robustness, preference precision and completeness, and information updating; scores are z-score normalized.Source: §5.4 · p. 8, Figure 5 · p. 8, Figure 6 · p. 8

6 Conclusion

The paper contributes a synthetic private-data generation pipeline and a benchmark for evaluating how well AI models understand personal information. Its findings suggest current methods are oversimplified for real-world scenarios involving noisy, fragmented, and irrelevant information.

6 Conclusion

  • PersonaBench provides a synthetic data generation pipeline and benchmark for evaluating AI models’ understanding of personal information from private data.Source: §6 · p. 8
  • The findings suggest current personalization methods are oversimplified for real-world scenarios involving noisy, fragmented, and irrelevant user information.Source: §6 · p. 8
  • The paper calls for more sophisticated methods and system designs to extract and use personal information from heterogeneous, imperfect private data.Source: §6 · p. 8

7 Limitation

The authors identify restricted release of ground-truth profiles, potentially offensive synthetic content, and occasional inconsistencies or unrealistic details as limitations.

7 Limitation

  • The underlying source-of-truth profiles and generation template will not be released, although generated evaluation documents are planned for open-source release.Source: §7 · p. 9
  • Some synthetic content may be perceived as offensive despite efforts to ensure realism and minimize harm.Source: §7 · p. 9
  • Automatically generated documents may contain occasional inconsistencies or unrealistic details, and perfect coherence cannot be guaranteed.Source: §7 · p. 9

8 *Ethical Statement

PersonaBench is an evaluation dataset made entirely from synthetic information, not real personal data or depictions of real-world individuals or patterns. The authors restrict its intended use to research and evaluation and advise attention to content sensitivities.

8 *Ethical Statement

  • PersonaBench is designed for evaluation and contains no real personal data; its synthetic profiles and conversations do not reflect real individuals, behaviors, or societal patterns.Source: §8 · p. 9
  • The dataset is intended strictly for research use, not commercial applications or decision-making systems affecting individuals or communities.Source: §8 · p. 9
  • Some generated content may be perceived as offensive or sensitive, so users are encouraged to apply additional content filtering when necessary and uphold responsible data handling.Source: §8 · p. 9

A.1 User Profile Example

The example profiles organize personal attributes into demographic, psychographic, and social categories, with attributes aligned to social relationships.

A.1 User Profile Example

  • Example profiles group attributes into demographic, psychographic, and social categories, aligning personal attributes with the characters’ social relationships.Source: §A.1 · p. 10

A.2 Social Graph Details

The appendix describes the persona descriptions and gives high-level prompt illustrations and example documents used in dataset generation. The examples show personal attributes appearing in conversation, user-AI interaction, and purchase-history documents.

A.2 Social Graph Details

  • Persona descriptions combine three PersonaHub samples with GPT-4o-generated personas, spanning roles from students and photographers to nurses and AI ethicists.Source: §A.2 · p. 10

A.3 High-level Prompt Illustration for Dataset Generation

A.4 Document Example

Loading 2502.20616v2…