Organizations: Columbia University, New York, NY, USA · St. Margaret’s Episcopal School, San Juan Capistrano, CA, USA · Carnegie Mellon University, Pittsburgh, PA, USA · Worcester Polytechnic Institute, Worcester, MA, USA · Institute for Social Research, University of Michigan, Ann Arbor, MI, USA · California State University Dominguez Hills, Carson, CA, USA
Mindfulness meditation is a widely accessible and evidence-based method for supporting mental health. Despite the proliferation of mindfulness meditation apps, sustaining user engagement remains a persistent challenge. Personalizing the meditation experience is a promising strategy to improve engagement, but it often requires costly and unscalable manual effort. We present MindfulAgents, a multi-agent system powered by large language models that (1) generates guided meditation scripts based on an expert-established mindfulness framework, (2) encourages users' reflection on emotional states and mindfulness skills, and (3) enables real-time personalization of the mindfulness meditation experience for each user. In a formative lab study (N=13), MindfulAgents significantly improved in-session engagement (p = 0.011) and self-awareness (p = 0.014), and reduced momentary stress (p = 0.020). Furthermore, a four-week deployment study (N=62) demonstrated a notable increase in long-term engagement (p = 0.002) and level of mindfulness (p = 0.023). Participants reported that MindfulAgents offered more relevant meditation sessions personalized to individual needs in various contexts, supporting sustained practice. Our findings highlight the potential of LLM-driven personalization for enhancing user engagement in digital mindfulness meditation interventions.
Figures & tables
Figure 1. MindfulAgents, a multi-agent system that provides expert-aligned, engaging, and personalized mindfulness meditation experience. The multi-agent system consists of: (A) an Expert-Alignment Agent that utilizes expert knowledge and feedback in the creation of safety meditation templates. (B) a Reflection Agent that encourages users to ponder their current emotional state, past meditation experiences, and technique-related questions, and (C) a Personalization Agent that builds on both the safety templates and reflective inputs to generate sessions that are user-specific and resonant. Figure 1 illustrates the MindfulAgents system, a multi-agent framework for expert-aligned, engaging, and personalized mindfulness meditation. On the left, inputs include expert knowledge and feedback, session preferences (mood, goal, duration), meditation techniques, and user profile data (personal info, historical sessions). At the top, the Expert-Alignment Agent produces safety meditation templates through collaboration with experts. At the bottom, the Reflection Agent engages users in interactive reflection based on present, past, and learning contexts. At the center, the Personalization Agent integrates safety templates, user profile, and reflection history to generate personalized mindfulness meditations. On the right, users receive personalized sessions via a mobile app, with feedback loops informing further personalization.
Figure 2. LLM Training and Retrieval Pipelines for Each Agent. Left) For the Expert-Alignment Agent, mindfulness practice curriculum scripts and UM concepts are used for supervised finetuning (SFT). Experts’ edits and feedback on the intermediate generated scripts are included as additional data to further tune the model via both SFT and direct preference optimization (DPO); Middle) For the Reflection Agent, meditation Q&A dialogues between instructors and students are used for SFT to emulate reflective conversations. And UM concepts are used for retrieval-augmented generation (RAG); Right) For the Personalization Agent, the SFT and DPO pipeline is similar to that of the Expert-Alignment Agent, except that standard curriculum scripts and mindfulness concepts are no longer included to better focus on goal-oriented scripts during the finetuning. Meanwhile, UM concepts are used for RAG to ensure content reliability. Figure 2 illustrates the LLM training and retrieval pipelines for three agents in the MindfulAgents system. On the left, the Expert-Alignment Agent is trained with mindfulness practice curriculum scripts using supervised finetuning (SFT), with iterative expert review and feedback loops applied through direct preference optimization (DPO). In the center, the Reflection Agent is finetuned on Q&A dialogue data from meditation instructors and students, using retrieval-augmented generation (RAG) for reflective conversations, and further guided by expert feedback. On the right, the Personalization Agent follows a similar pipeline to the Expert-Alignment Agent, but excludes standard curriculum scripts during finetuning, instead leveraging goal-oriented scripts and UM concepts via RAG. Across all agents, expert-in-the-loop processes ensure alignment, reliability, and personalization of outputs.
Table 1. Ablation Studies Comparing Input Structure of the Personalization Agent. Average expert ratings across alignment, personalization, and overall quality (mean ± standard deviation across simulated persona). Scores of the same condition in two studies may vary as the content was re-generated in each ablation comparison. Wilcoxon signed-rank tests were conducted to compare the scores. r = rank-biserial correlation effect size. Statistical significance: p<0.05∗ , p<0.01∗∗ , p<0.001∗∗∗ .
Figure 3. The User Flow across MindfulAgents Interface. (A) The user starts by sharing their mood, goals, as well as preferred mindfulness technique, duration, and level of guidance. (B) The user then has the option to interact with the Reflection Agent to share their thoughts and experience. (C) The user could read the tip and personal summary cards while waiting for the personalized meditation content and audio being generated. (D) The user enters the meditation experience with an audio player and reports feedback at the end of the meditation session. Figure 3 shows the user flow across the MindfulAgents mobile interface. From left to right: (A) Users begin by sharing context, including mood, goals, meditation technique, duration, and guidance preference. (B) Before meditation, they can engage in self-reflection through a chatbot interface. (C) While waiting, users view personalized tips and summary cards for learning and reflection. (D) Finally, they participate in a personalized meditation session via an audio player and provide feedback afterward.
Figure 4. Overview of the Formative Lab Study and Field Deployment Study. (A) Formative Study used a within-subject design (N=13) to compare user experience across three ablation conditions: StaticAgent (MindfulAgents without personalization or reflection), PersonalAgents (MindfulAgents without reflection), and the full MindfulAgents system. (B) Deployment Study used a between-subjects design (N=62) to compare StaticAgent and MindfulAgents over a four-week in-the-wild deployment, focusing on long-term engagement and changes in mindfulness and broader well-being. Figure 4 shows study designs in flowchart format for two user studies of MindfulAgents. Shown at the top, (A) Formative Study: within-subject design with onboarding (30 min), three 10-minute conditions in counterbalanced order, and a 30-minute exit interview for qualitative feedback and condition comparison. Shown at the bottom, (B) Deployment Study: between-subject design with onboarding (30 min), a two-week UM Standard phase, pre-survey, two-week experimental condition (X or Y), post-survey, and a 30-minute exit interview. Measures include physical health (PSQI), mental well-being (PANAS, GAD-7), and mindfulness (FFMQ).
Figure 5. Formative Study Results. (A) Scores across five user experience metrics demonstrate that MindfulAgents has the overall best performance. Error bars indicate standard error. (B) Preference rankings show consistent findings. MindfulAgents received the highest ranking overall, followed by PersonalAgents, and then StaticAgent. Figure 5 presents results from the Formative Lab Study. On the left, (A) Bar chart of in-session user experience ratings (1–10) across five metrics: in-session engagement, level of personalization, deepen awareness, stress reduction, and willingness for future use. MindfulAgents show consistently higher scores than PersonalAgents and StaticAgent, with significant improvements in engagement, awareness, and stress reduction (p < 0.05). Error bars show standard error. On the right, (B) Stacked bar charts of preference rankings across overall preference, helpfulness, and potential for future use. MindfulAgents is most frequently ranked first, followed by PersonalAgents, then StaticAgent.
Figure 6. Deployment Study Results on Long-Term Engagement during the Free Practice Phase. (A) MindfulAgents achieved a higher post-intervention engagement than that of StaticAgent, and (B) this advantage sustains over time throughout the free practice period. Figure 6 compares post-intervention engagement between StaticAgent and MindfulAgents. On the left, (A) Bar chart shows mean daily sessions during the free meditation stage: MindfulAgents users practiced significantly more (0.94 ± 0.07) than StaticAgent users (0.40 ± 0.04). On the right, (B) Line graph shows sessions per day across 14 days: MindfulAgents maintained consistently higher engagement, with rates increasing over time, while StaticAgent remained low and flat.
Figure 7. Deployment Study Results on the Changes in Behavioral Outcomes. Scores were computed as the delta between Post- and Pre-surveys. (A) mindfulness (FFMQ-SF), (B) mood (PANAS-SF), (C) anxiety (GAD-7), and (D) sleep quality (PSQI). MindfulAgents shows better results in promoting mindfulness and mood level compared to the StaticAgent. Figure 7 shows changes in behavioral outcomes, computed as post–pre survey deltas, comparing StaticAgent and MindfulAgents. (A) Mindfulness (FFMQ-SF): MindfulAgents produced significantly greater improvement than StaticAgent, with mean gains around +4. (B) Mood (PANAS-SF): MindfulAgents users showed significant within-condition improvement, though not significantly different from StaticAgent. (C) Anxiety (GAD-7): Both groups showed small, nonsignificant changes. (D) Sleep quality (PSQI): Minimal changes in both conditions. Significance markers indicate statistical tests, with only mindfulness showing between-condition differences favoring MindfulAgents.
Figure 8. Deployment Study Results on Final User Satisfaction Scores. MindfulAgents shows overall higher scores across all dimensions on user experience and content quality, although no statistical significance was observed. Figure 8 shows post-study participant satisfaction ratings of StaticAgent and MindfulAgents across two dimensions: User Experience (goal alignment, inclusiveness) and Content Quality (guidance and pace, concept and clarity). In all four categories, MindfulAgents received higher mean ratings than StaticAgent, with differences most pronounced for goal alignment and guidance and pace, although Mann-Whitney U tests did not show significance. Error bars indicate standard error.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 9. Overview of the iterative system design. Each panel shows one of eight design iterations, from a vanilla GPT–4o baseline to the final MindfulAgents architecture integrating Expert-alignment, Personalization, and Reflection agents. A multi-panel diagram showing eight successive system iterations. Early panels depict a single LLM meditation coach with unstable outputs. Middle panels add finetuning and format checks to reduce hallucinations and enforce script structure. Later panels introduce separate agents for expert-aligned template creation, real-time personalization, and user reflection. The final panel shows the complete MindfulAgents pipeline combining expert templates, reflective chat, and personalized meditation generation.
Iteration
Configuration
Key observations
1
Vanilla GPT-4o
Vanilla GPT-4o prompted as a meditation coach. Extremely unstable: frequent hallucinations, wrong format for the audio API, and inconsistent adherence to the Unified Mindfulness curriculum or expert guidance.
2
Expert-aligned finetuned model
GPT-4o finetuned on expert-written UM scripts. Reduced some obvious errors, but the model still hallucinated and occasionally produced responses in an incorrect format for the audio API .
3
Finetuning + format checks
Added evaluation layers (script length, script format, and script ending) to validate and correct outputs before passing them to the audio API. Reliability improved substantially, but the model still mishandled fine-grained details of the UM curriculum .
4
Expert-alignment agent
Restructured the system as an expert-aligned agent with additional curriculum alignment in evaluation and prompting. Eliminated most formatting and curriculum-consistency issues, but the multi-step agent pipeline introduced high latency and increased risk of failure during long generations , motivating extraction and caching of expert-aligned scripts.
5
Expert-alignment agent + personalization LLM
Introduced a dedicated personalization LLM to adapt expert-aligned scripts to user-specific factors (e.g., core personalization inputs identified in DG 3) using faster deterministic checks. Improved responsiveness and personalization, but revealed the need for stronger safety resources and clearer escalation paths for sensitive disclosures .
6
Expert-alignment agent + personalization agent
Expanded the personalization component into a full personalization agent with appropriate safety guardrails. Expert feedback highlighted the need for richer user context and a stronger sense of therapeutic alliance .
Appendix
Table 2. Overview of design iterations leading to the final MindfulAgents system.
UM Technique
Description
See-Hear-Feel
A systematic way Unified Mindfulness organizes sensory experience into three categories: sight, sound, and body sensation. “See” includes outer sights of the world and inner visual thinking on the mental screen. “Hear” includes outer sounds of the world and inner auditory thinking, such as self-talk. “Feel” includes outer physical body sensations and inner emotional body sensations.
Restful States
Any sensory experience characterized by stillness, quiet, or relaxation, either arising naturally or intentionally created. Includes visual rest such as light, darkness, or absence of images, auditory rest such as silence or a quiet mind, and physical or emotional rest such as relaxation, neutrality, or peace. Restful states support tranquility, help induce equanimity, and create a positive feedback loop that sustains motivation to practice.
Gone
The instant you become aware that an experience is partially or completely ending, whether you’re clear about what the experience is or not.
Concentration Power
One of the three core attention skills, the ability to focus on what you choose when you want to. You can think of it as a kind of flexible focus. Can be developed by focusing on a small or large object of focus, and for a short or long period of time. Can be developed by keeping your focus within a focus range for as long as possible (sustained concentration), or by focusing for brief moments on one or more sensory experiences within a focus range (momentary concentration).
Sensory Clarity
One of the three core attention skills, the ability to track and explore sensory experience in real time. Has two sides: Detection (sensitivity to details) and Discernment (the ability to tell this from that, such as the distinction between See and Hear and Feel).
Equanimity
One of the three core attention skills, the ability to allow sensory experience to come and go without push and pull. Can be developed through a range of methods, such as noticing when it happens spontaneously, dropping judgments, intentionally relaxing, non-interference, anchoring attention away, noticing disappearances, creating or collapsing distance between yourself and a focus object (through zooming, being the witness, etc.), or accessing a positive emotional state, such as gratitude, compassion, etc.
Appendix
Table 3. List of Unified Mindfulness (UM) techniques.
Category
Goal
Starting Day
Positivity, Express Yourself, Better Habits, Find Your Purpose, Baseline Happiness
Ready to Work
Improve Focus, Peak Performance, Be More Present, Creative Work
Taking a Break
Work Break, Rejuvenation, Gain Clarity
SOS
Panic Attack, Pain, Challenging Situations
Socializing
Support Others, Emotional Intimacy, Manage Conflict, Enjoy Social Life, Forgiveness
Big Event
Public Speaking, Test / Interview, Difficult Conversation
Appendix
Table 4. List of category and goal pairs for goal-oriented practice scripts.
ID
Name
Persona Description
Goal
1
Sarah
Sarah is a 34-year-old marketing director at a Fortune 500 company in Chicago. She works 60+ hour weeks, manages a team of 15 people, and frequently travels for business. Recently divorced with joint custody of her 8-year-old daughter, Sarah struggles to balance her demanding career with single motherhood. She experiences chronic stress, has trouble sleeping, and often feels like she’s constantly running on empty.
Panic Attack (Study 1), Work Break (Study 2), Forgiveness and Letting Go (Study 3), Sleep (Study 4)
2
Marcus
Marcus is a 20-year-old pre-med student at a competitive university. He’s the first in his family to attend college and feels immense pressure to succeed academically while working part-time to help support his family financially. Marcus experiences significant social anxiety, especially in group settings and during presentations. He has panic attacks during exams and often feels imposter syndrome among his peers. Having grown up in a traditional Latino household, he’s somewhat skeptical about meditation but is desperate for relief from his constant worry and self-doubt. Marcus wants to find ways to calm his mind before tests, build confidence in social situations, and develop better emotional regulation skills.
Test / Interview (Study 1), Enjoy Social Life (Study 2), Improve Focus (Study 3), Test / Interview (Study 4)
3
Ellie
Ellie is a 67-year-old retired art teacher who has been exploring various spiritual practices since her husband passed away three years ago. She lives alone in a small coastal town in Maine and has plenty of time for contemplative practices. Ellie has tried yoga, attended meditation retreats, and reads extensively about mindfulness and Eastern philosophy. However, she often feels like she’s "doing it wrong" and gets frustrated when her mind wanders during meditation. She’s looking for deeper spiritual connection and meaning in her life, wanting to move beyond surface-level relaxation techniques.
Find Your Purpose (Study 1), Surprise Me (Study 2), Create More Positive Situations (Study 3), Forgiveness and Letting Go (Study 4)
4
David
David is a 28-year-old personal trainer and former Division I soccer player who lives in Austin, Texas. He’s always been focused on physical performance and views his body as a machine to be optimized. Recently, David has been dealing with chronic pain from old sports injuries and has been recommended meditation by his physical therapist as part of his recovery. He’s naturally skeptical of anything that seems "touchy-feely" or unscientific, preferring evidence-based approaches. David is also struggling with anger management issues that have affected his relationships and coaching style with clients.
Priya is a 45-year-old nurse practitioner who works in a busy emergency department in Phoenix. She’s been on the frontlines of healthcare for over 15 years and has experienced significant burnout, especially after the pandemic. Priya is also the primary caregiver for her elderly mother who has dementia, while raising two teenage children with her husband. She feels emotionally drained, has developed compassion fatigue, and struggles with guilt about not being able to give her best to her patients, family, or herself. Priya has some familiarity with meditation through her Hindu cultural background but hasn’t practiced regularly since childhood.
Support Others (Study 1), Deep Relaxation (Study 2), Gain Psychological Clarity (Study 3), Rebound from Challenging Situations (Study 4)
6
Jimmy
Jimmy is a 31-year-old graphic designer and freelance artist living in Brooklyn, New York. He quit his corporate design job two years ago to pursue his passion for illustration and has been struggling financially ever since. Jimmy works irregular hours, often pulling all-nighters to meet client deadlines, and his income is unpredictable. He battles creative blocks, imposter syndrome, and the constant pressure to find new clients while maintaining his artistic integrity. Jimmy has experimented with recreational substances to enhance creativity but realizes this isn’t sustainable. He’s naturally introspective and has tried meditation apps sporadically but finds it hard to sit still when his mind is racing with creative ideas or financial worries.
Creative Work (Study 1), Express Yourself (Study 2), Deep Relaxation (Study 3), Rejuvenation (Study 4)
Appendix
Table 5. List of ablation study persona and their profile information
ID
Name
Persona Description
Goal
7
Becca
Becca is a 52-year-old administrative assistant from suburban Minneapolis whose youngest child just left for college. After 25 years of marriage, she and her husband are getting divorced, and she’s living alone for the first time in her adult life. Becca has always defined herself through her roles as wife and mother, and now feels lost and unsure of her identity. She’s dealing with perimenopause symptoms, including mood swings and sleep disturbances, and has gained weight that affects her self-esteem. Becca has never meditated before but has been recommended mindfulness by her therapist to help with the life transition.
Michael is a 38-year-old associate professor of economics at a prestigious university who is up for tenure review next year. He’s published extensively but constantly worries that his research isn’t groundbreaking enough. Michael works 70+ hour weeks, has chronic insomnia, and experiences frequent tension headaches from stress. He’s highly analytical and approaches everything, including his personal life, with rigorous logical thinking. Michael has read about meditation’s cognitive benefits in scientific journals but has never tried it because he considers it "unscientific" and worries about losing his analytical edge. He’s recently been having panic attacks before important presentations and realizes he needs help.
Public Speaking (Study 1), Improve Focus (Study 2), Express Yourself (Study 3), Public Speaking (Study 4)
9
Carmen
Carmen is a 24-year-old lifestyle influencer from Los Angeles with 200K followers across platforms. She built her brand around wellness and positivity but privately struggles with anxiety, depression, and the pressure to constantly perform happiness online. Carmen experiences severe FOMO, compares herself constantly to other influencers, and has developed an unhealthy relationship with social media metrics. She’s tried various wellness trends for content but has never committed to a consistent meditation practice. Carmen feels like a fraud promoting wellness while struggling internally and fears that being authentic about her mental health struggles would damage her brand.
Gain Clarity (Study 1), Baseline Happiness (Study 2), Experience Closeness (Study 3), Enjoy Social Life (Study 4)
10
Bob
Bob is a 66-year-old former construction foreman from rural Montana who retired six months ago after 40 years in the trades. He’s struggling with the transition from a physically demanding, structured work life to having unlimited free time. Bob feels useless and bored, misses the camaraderie of his work crew, and is experiencing mild depression. His wife suggests he needs new hobbies, but Bob feels too old to learn new things and worries about his declining physical abilities. He’s skeptical of anything that seems "new age" or feminine, viewing meditation as something for "yoga people," not tough guys like himself. Bob has chronic lower back pain from years of physical labor and occasional anger outbursts that worry his family.
Aisha is a 26-year-old PhD candidate in psychology at a competitive research university. She’s in her fourth year of graduate school, working on her dissertation while teaching undergraduate courses and conducting research. Aisha feels overwhelmed by academic pressure, imposter syndrome, and the uncertain job market in academia. She’s passionate about her research on trauma therapy but finds herself emotionally affected by the heavy subject matter. Aisha comes from a working-class family and is the first to pursue higher education, feeling pressure to succeed not just for herself but to prove that her family’s sacrifices were worthwhile. She has some experience with mindfulness through her psychology training but has never developed a personal practice.
Support Others (Study 1), Rejuvenation (Study 2), Work Break (Study 3), Support Others (Study 4)
12
Tony
Tony is a 29-year-old security guard who works the overnight shift at a corporate office building in Seattle. He’s been working nights for three years to earn extra money while his wife attends nursing school during the day. Tony struggles with maintaining a healthy sleep schedule, often feeling groggy and disconnected from his family and friends who are on normal daytime schedules. He feels isolated and lonely during his shifts, often mindlessly scrolling social media to pass time. Tony has gained weight from irregular eating patterns and lack of exercise, and he’s become increasingly irritable with his wife and young daughter during the few hours they’re all awake together. He’s never tried meditation but is desperate for something to help him feel more balanced.
Table 5. List of ablation study persona and their profile information (cont.)
Metric
Questionnaire item
Explanation
In-session engagement
“The session felt engaging.”
Assesses how immersed, attentive, and focused participants felt during the meditation session.
Perceived level of personalization
“It addressed my current mood/goals.”
Captures the extent to which participants felt the guidance was tailored to their present emotional state and intentions.
Deepening self-awareness
“It helped deepen my awareness.”
Indicates whether the session supported noticing and understanding inner experiences, thoughts, and bodily sensations.
Willingness for future use
“I would choose to use this again.”
Reflects participants’ intention to use this type of session in the future, as a proxy for perceived long-term value.
Stress reduction
“My stress level reduced.”
Measures perceived short-term relief from stress or tension following the session.
Appendix
Table 6. Single-item measures used in the Formative Lab Study (Study 1). All items were rated on a 5-point Likert scale (1 = strongly disagree, 5 = strongly agree).
Digital mental health (DMH) tools have extensively explored personalization of interventions to users' needs and contexts. However, this personalization often targets what support is provided, not how it is experienced. Even well-matched content can fail when the interaction format misaligns with how someone can engage. We introduce generative experience as an approach to DMH support, where the intervention experience is composed at runtime. We instantiate this in GUIDE, a system that generates personalized intervention content and multimodal interaction structure through rubric-guided generation of modular components. In a preregistered study with N=237 participants, GUIDE significantly reduced stress (p=.02) and improved user experience (p=.04) compared to an LLM-based cognitive restructuring control. GUIDE also supported diverse forms of reflection and action through varied interaction flows, while revealing tensions around personalization across the interaction sequence. This work lays the foundation for interventions that dynamically shape how support is experienced and enacted in digital settings.
Ananya Bhattacharjee, Michael Liut, Matthew Jörke +2
Stanford University, Stanford, California, USA · University of Toronto Mississauga, Mississauga, Ontario, Canada
We present Tell Me, a mental well-being system that leverages advances in large language models to provide accessible, context-aware support for users and researchers. The system integrates three components: (i) a retrieval-augmented generation (RAG) assistant for personalized, knowledge-grounded dialogue; (ii) a synthetic client-therapist dialogue generator conditioned on client profiles to facilitate research on therapeutic language and data augmentation; and (iii) a Well-being AI crew, implemented with CrewAI, that produces weekly self-care plans and guided meditation audio. The system is designed as a reflective space for emotional processing rather than a substitute for professional therapy. It illustrates how conversational assistants can lower barriers to support, complement existing care, and broaden access to mental health resources. To address the shortage of confidential therapeutic data, we introduce synthetic client-therapist dialogue generation conditioned on client profiles. Finally, the planner demonstrates an innovative agentic workflow for dynamically adaptive, personalized self-care, bridging the limitations of static well-being tools. We describe the architecture, demonstrate its functionalities, and report evaluation of the RAG assistant in curated well-being scenarios using both automatic LLM-based judgments and a human-user study. This work highlights opportunities for interdisciplinary collaboration between NLP researchers and mental health professionals to advance responsible innovation in human-AI interaction for well-being.
Large language models have enabled agentic systems that reason, plan, and interact with tools and environments to accomplish complex tasks. As these agents operate over extended interaction horizons, their effectiveness increasingly depends on adapting behavior to individual users and maintaining continuity across interactions, giving rise to personalized LLM-powered agents (PLAs). In such long-term, user-dependent settings, personalization permeates the entire decision pipeline rather than remaining confined to surface-level response generation. This survey provides a capability-oriented review of personalized LLM-powered agents. Existing work is organized around four interdependent capabilities: profile modeling, memory, planning, and action execution. Using this taxonomy, representative methods are synthesized and analyzed to illustrate how user signals are represented, propagated, and utilized across the agent pipeline, highlighting cross-component interactions and recurring design challenges. Evaluation metrics and benchmarking paradigms tailored to personalized agents are further examined, along with application scenarios ranging from conversational assistants to domain-specific expert systems. By clarifying the design space of personalization in agent systems, this survey provides a structured foundation for developing more user-aligned, adaptive, and deployable LLM-powered agents.
Yue Xu, Qian Chen, Zizhan Ma +5
ShanghaiTech University, China · Tongji University, China · The Chinese University of Hong Kong, China +3