PersonaTrace: Synthesizing Realistic Digital Footprints with LLM Agents

Minjia Wang; Yunfeng Wang; Xiao Ma; Dexin Lv; Qifan Guo; Lynn Zheng; Benliang Wang; Lei Wang

cs.CLMar 12, 2026

PersonaTrace: Synthesizing Realistic Digital Footprints with LLM Agents

Authors: Minjia Wang, Yunfeng Wang, Xiao Ma, Dexin Lv, Qifan Guo, Lynn Zheng, Benliang Wang, Lei Wang,

Organizations: Apple · Harvard University

Abstract

Digital footprints (records of individuals' interactions with digital systems) are essential for studying behavior, developing personalized applications, and training machine learning models. However, research in this area is often hindered by the scarcity of diverse and accessible data. To address this limitation, we propose a novel method for synthesizing realistic digital footprints using large language model (LLM) agents. Starting from a structured user profile, our approach generates diverse and plausible sequences of user events, ultimately producing corresponding digital artifacts such as emails, messages, calendar entries, reminders, etc. Intrinsic evaluation results demonstrate that the generated dataset is more diverse and realistic than existing baselines. Moreover, models fine-tuned on our synthetic data outperform those trained on other synthetic datasets when evaluated on real-world out-of-distribution tasks.

Figures & tables

Figure 1: An overview of our methods. The Persona Agent generates a basic profile from a demographic prior, and iteratively adding realistic attributes to it. The Event Agent retrieves seed events from the event memory and aligns them to the persona, and brainstorms with self-reflection to generate an event forest that serves as the scaffolding of the digital footprints. The Artifact Generator Agent and Critic Agents forms a loop - the Artifact Generator Agent generates the outline and then digital artifacts, and the critic agents provides feedback to iteratively improve the quality of the artifacts.

Dataset		Pairwise Corr. (↓)	Remote-Clique (↑)	Entropy (↑)	Avg. #Links	Avg. Length
	FinePersonas-Email	0.2333	0.7680	2.8080	0.0000	681.50
	IWSPA-2023-Adversarial	0.4079	0.5919	2.5094	0.0000	276.48
	LLM-Gen Phishing	0.3094	0.6906	2.6969	0.2522	685.58
	Synthetic-Satellite-Emails	0.5416	0.4586	2.8218	0.0000	1050.76
Synthetic	PersonaTrace (Proposed)	0.2093	0.7898	2.8305	0.2532	1437.87
	Enron	0.2798	0.7218	2.7110	0.0000	2002.57

Table 1: Diversity and realism of datasets related to emails. Best results in synthetic datasets are in bold.

Dataset		Tone	Fluency	Coherence	Informativeness	Engagement	Overall
Synthetic	FinePersonas-Email	4.52	4.92	4.56	3.98	3.98	4.39
	IWSPA-2023-Adversarial	1.59	1.91	1.67	1.31	1.22	1.53
	LLM-GenPhishing	3.41	4.89	4.67	3.13	2.78	3.70
	Synthetic-Satellite-Emails	4.82	4.96	4.86	4.91	3.65	4.64
	PersonaTrace (Proposed)	4.95	4.99	4.99	4.92	4.09	4.79
	Enron	4.19	4.73	4.47	4.28	2.71	4.07

Table 2: LLM-As-Judge scores of datasets related to emails. Best results in synthetic datasets are in bold.

Email Categorization
Training Set	Enron		Human-Gen Phishing		Private		Private w/o Spam		W3C-Emails
	Acc	F1	Acc	F1	Acc	F1	Acc	F1	Acc	F1
None	0.2741	0.0722	0.0818	0.0236	0.0011	0.0022	0.0025	0.0028	0.0011	0.0004
FinePersonas-Email	0.5908	0.1810	0.2241	0.0755	0.0015	0.0027	0.0020	0.0022	0.5401	0.1093
IWSPA-2023-Adversarial	0.0010	0.0038	0.0007	0.0004	0.0012	0.0023	0.0035	0.0039	0.0011	0.0022
LLM-Gen Phishing	0.4046	0.1095	0.1363	0.0376	0.0008	0.0016	0.0015	0.0017	0.1909	0.0523

Table 3: Extrinsic evaluation results on email-related downstream tasks: email categorization, email drafting, and question answering. Best results are in bold.

Figure 2: UMAP visualization of the dataset embeddings related to text messages and conversations. Synthetic datasets are denoted by circles, while real datasets are represented by triangles. Among the synthetic datasets, PersonaTrace appears closest in the embedding space to the private dataset and u-sticker, indicating a higher degree of realism and alignment with real-world digital communications.

Training Set	DailyDialog		Persona-Chat		Private		u-sticker
	Acc	BertS	Acc	BertS	Acc	BertS	Acc	BertS
None	0.1202	0.0237	0.1072	0.1521	0.0958	0.1102	0.0933	0.0849
DailyConversations	0.1091	0.1867	0.1070	0.2195	0.0961	0.2576	0.0933	0.2976
DiaSynth	0.1182	0.1334	0.1089	0.2036	0.0961	0.1698	0.0875	0.1666
PersonaBench	0.1253	0.0328	0.1064	0.1663	0.0960	0.1246	0.0930	0.0986
Synthetic-Persona-Chat	0.1101	0.2143	0.0982	0.2272	0.0955	0.2954	0.0893	0.3306

Table 4: Extrinsic evaluation results on message-related downstream task: next message prediction. Best results are in bold.

Task		w/o Agents	w/ Agents
	Acc	0.0063	0.2733
Email Categorization	F1	0.0057	0.0813
	ROUGE	0.0376	0.1527
Email Drafting	BertS	0.3012	0.4590
	ROUGE	0.0413	0.2500
Question Answering	BertS	0.2880	0.4405

Table 5: Comparison of downstream task performance between the agent-ablated and full implementations.

Figure 3: Comparison of diversity and realism between the agent-ablated and full implementations. For Pairwise Correlation, lower values indicate greater diversity. For Remote-Clique and Entropy, higher values reflect greater diversity. Average number of links per email and average email length are used as indicators of realism, with higher values suggesting closer resemblance to human-generated emails.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Prompts for LLM-As-judge evaluation.

Panel of Figure 5 — Figure 5: Prompt for generating comprehensive and culturally grounded profile.

Panel of Figure 6 — Figure 6: Prompt for generating comprehensive and culturally grounded profile.

Panel of Figure 7 — Figure 7: Prompt for brainstorming comprehensive, profile-based events.

Panel of Figure 8 — Figure 8: Prompt for generating related event expansions from a single event description.

Panel of Figure 9 — Figure 9: Prompt for generating structured email body outlines based on event data.

Panel of Figure 10 — Figure 10: Prompt for generating realistic emails from outlines and event references.

Panel of Figure 11 — Figure 11: Prompt for expert-level email review and constructive feedback.

Panel of Figure 12 — Figure 12: Prompt for revising and improving emails based on specific feedback.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 625 papers

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent
Jun 24, 2026Sriram Selvam, Anneswa GhoshSynthetic Data Generation Synthetic Benchmark Generation
SYN-DIGITS: A Synthetic Control Framework for Calibrated Digital Twin Simulation
Apr 8, 2026Grace Jiarui Fan, Chengpiao Huang, Tianyi Peng +2Synthetic Control Language Model Calibration
Finance Division, Columbia Business School. · Department of IEOR, Columbia University · Decision, Risk, and Operations Division, Columbia Business School. +1
Private Seeds, Public LLMs: Realistic and Privacy-Preserving Synthetic Data Generation
Apr 8, 2026Qian Ma, Sarah RajtmajerPrivacy-Preserving Language Models Synthetic Data Generation
Information Sciences and Technology, The Pennsylvania State University

Jun 24, 2026J/K move · Enter open · S save

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent

Sriram Selvam, Anneswa Ghosh

Synthetic Data Generation Synthetic Benchmark Generation User Modeling Synthetic Data Evaluation Synthetic Data Generation for LLM Agents Synthetic Population Generation

Open paper AskPDF arXiv

Foundation-model research increasingly needs data about people: user state, personal histories, relationships, contact-like fields, documents, and longitudinal updates. Real user data is difficult to share, perturb, audit, or redistribute responsibly, while independently generated fake fields rarely preserve the cross-field and temporal consistency needed for controlled evaluation. We present PROFILEFOUNDRY, a deterministic generator and fixed reference release of 100,000 adult synthetic Person Objects across eight locales. Each object combines a typed current snapshot, household, family, and employer links, snapshot-aligned events, normalized relational views, and generation provenance. The release contains 709,228 events, 40,338 households, 52,491 employers, and 518,564 directed relationship edges. We report evidence in separate categories: selected population-marginal comparisons, per-object invariant checks, release-wide referential and temporal closure, and coincidence/provenance screens. PROFILEFOUNDRY is not a population-fidelity model, a rendered-text corpus, or a formal privacy mechanism. Instead, it is a responsible synthetic source layer for constructing downstream foundation-model evaluations involving memory, privacy, document understanding, record linkage, and agent state while keeping the synthetic person behind each artifact inspectable

Abstract

Synthetic Data Generation

Explore similar work

ProfileFoundry: A Synthetic Person-Object Substrate for Privacy, Memory, and Tool-Use Evaluation in LLM Agent

SYN-DIGITS: A Synthetic Control Framework for Calibrated Digital Twin Simulation

Private Seeds, Public LLMs: Realistic and Privacy-Preserving Synthetic Data Generation