UbiComp/ISWC 2026 Tutorial | LLM Personalization

The Human Question

When should a system adapt—and when should it step back?

Large language models are moving into phones, wearables, vehicles, homes, and public environments. In these settings, useful personalization must respond to people, places, routines, and changing intent—not only to a static profile.

This tutorial connects personalization methods with cloud–edge–device systems. We examine what stays close to the person, what can be coordinated at the edge, and when cloud intelligence is worth the latency and disclosure.

Our central premise is simple: people should be able to inspect, correct, pause, and revoke adaptation. Success therefore includes agency, privacy, fairness, latency, energy, and longitudinal benefit—not accuracy alone.

Legible Correctable Revocable Context-aware

Why This Tutorial

From model-centric personalization to lived experience

01

Situated, not static

Model evolving goals, social settings, device handovers, and moments when personalization is inappropriate.

02

Collaborative by design

Place memory, adaptation, and inference across device, edge, and cloud under real resource constraints.

03

People retain agency

Build consent, explanation, correction, expiration, deletion, and a non-personalized fallback into the loop.

04

Evaluated in the world

Connect model and systems metrics with workload, interruption, trust, equity, and longitudinal outcomes.

Design Space

One person, three computational layers

Placement is a human decision as much as a systems decision: it determines what can be known, how quickly the system responds, and which boundaries remain meaningful.

Device

Close to the person

Private context, recent memory, sensors, lightweight adapters, immediate feedback, offline continuity.

Lowest latency
Smallest resource budget
Strongest local control

Edge

Close to the situation

Local coordination, caching, cross-device context, adaptive routing, and low-latency shared services.

Contextual orchestration
Selective aggregation
Trust must be negotiated

Cloud

Broad capability

General knowledge, high-capacity reasoning, expensive updates, and support for smaller local models.

Greatest model capacity
Network-dependent
Highest disclosure cost

Human agency is the routing policy Inspect · Correct · Pause · Revoke

October 11 · 14:00–17:00

Six parts, one personalization journey

The afternoon program moves from foundations through personalized prompting, adaptation, and alignment, then connects these methods to cloud–edge–device collaboration before a concise synthesis and Q&A.

Time	Module	Presenter
14:00–14:20	Part I – Introduction20 min	Ruijie Wang
14:20–14:50	Part II – Personalized Prompting30 min	Ruijie Wang
14:50–15:30	Part III – Personalized Adaptation40 min	Qingkai Zeng
15:30–16:00	Coffee Break30 min	—
16:00–16:20	Part IV – Personalized Alignment20 min	Xin Hui
16:20–16:40	Part V – Cloud–Edge–Device Collaboration20 min	Xin Hui
16:40–16:50	Part VI – Summary & Looking Forward10 min	Ruijie Wang
16:50–17:00	Q&A10 min	All speakers

Organizers & Instructors

A cross-disciplinary teaching team

Expertise spanning personalized LLMs, efficient AI systems, user modeling, trustworthy learning, and large-scale data mining.

Ruijie Wang

Beihang University

Professor working on trustworthy and efficient AI, large language models, graph learning, temporal data mining, and structured knowledge integration. His work connects personalization methods with efficient and responsible deployment.

Homepage

Qingkai Zeng

Nankai University

Associate Professor researching text mining, knowledge-enhanced LLMs, and reliable and interpretable reasoning in knowledge-intensive systems.

Homepage

Xin Hui

Xi'an Jiaotong University

Leads Parts IV and V on personalized alignment and cloud–edge–device collaboration.

Xuefei Wang

Beihang University

Ph.D. student focused on LLM personalization and system implementation, with hands-on experience in personalization pipelines and deployment trade-offs.

Yuhan Wang

Beihang University

Incoming Ph.D. student researching LLM personalization and continual learning, with a focus on evolving user preferences and continual on-device adaptation.

Li Sun

Beijing University of Posts and Telecommunications

Associate Professor researching large language models, graph representation learning, knowledge mining, and efficient knowledge-enhanced adaptation.

Homepage

Ge Wang

Xi'an Jiaotong University

Tutorial organizer and chair supporting the program on personalized AI.

Shengzhong Liu

Shanghai Jiao Tong University

Tutorial organizer and chair supporting the program on personalized AI.

Jizhong Zhao

Xi'an Jiaotong University

Tutorial organizer and chair supporting the program on personalized AI.

Jianxin Li

Beihang University

Professor working on social network analysis, trustworthy machine learning, large-scale data mining, and dependable learning systems.

Homepage

Philip S. Yu

University of Illinois Chicago

Distinguished Professor and Wexler Chair whose research spans data mining, graph learning, user behavior modeling, multimodal LLMs, and large-scale analytics.

Homepage

Open Materials

Use the framework after Shanghai

The proposal is available now. Slides, lecture notes, architecture comparison tables, and the reading map will be released here.

PDFTutorial BrochureDownload the final program and presenter overview

QR code for the UbiComp 2026 tutorial website

WebsiteOfficial Tutorial WebsiteScan the QR code or open the program online PDFTutorial ProposalRead the UbiComp/ISWC 2026 proposal

SlidesTeaching DeckComing soon

NotesLecture NotesComing soon

ReferenceArchitecture TablesComing soon

Citation

BibTeX

Tutorial website citation; update with the ACM Digital Library record when available.

@misc{wang2026human,
  title        = {From Generic Intelligence to Personalized AI: A Tutorial on Foundations of LLM Personalization},
  author       = {Wang, Ruijie and Zeng, Qingkai and Hui, Xin and Wang, Xuefei and Wang, Yuhan and Sun, Li and Wang, Ge and Liu, Shengzhong and Zhao, Jizhong and Li, Jianxin and Yu, Philip S.},
  howpublished = {Tutorial at ACM UbiComp/ISWC 2026},
  year         = {2026},
  address      = {Shanghai, China},
  url          = {https://llm-personalization-tutorial.github.io/ubicomp2026-tutorial/}
}

Contact

Questions, scenarios, collaborations

xuefeiw@buaa.edu.cn ruijiew@buaa.edu.cn Tutorial Website