Zhaoxuan Tan 「谭兆轩」

profile photo

How to pronounce my name?

Zhaoxuan -> jow sh-yen.
Tan -> tæn.
You can just call my last name Tan, which is much easier to pronounce.

I'm a final-year PhD candidate in Computer Science and Engineering at the University of Notre Dame, advised by Meng Jiang. I build personalized foundation models and agents that learn from what people write and how they interact.

My research asks how a model can infer someone’s preferences from their history and adapt to a new user without a separate round of training. I also study what agents should remember over time, both for one person and for a team whose members may remember the same conversation differently.

Before Notre Dame, I studied computer science at Xi’an Jiaotong University (2019–2023), where I worked with Shangbin Feng and Minnan Luo in the LUD lab.

I'm on the job market and seeking full-time Research Scientist or Applied Scientist positions in industry. If you think I might be a good fit, please feel free to reach out to me via email!

Email:  ztan3 [at] nd [dot] edu  /  tanzx9 [at] gmail [dot] com

CV  /  Google Scholar  /  Semantic Scholar  /  X (Twitter)  /  Github  /  LinkedIn

profile photo

Research Overview

Selected publications

When two people ask the same question, their histories can make different answers appropriate. My work explores what those histories reveal, how they can shape a model, and what an agent needs to remember when several people work together.

Reading preferences from behavior

What can we learn about someone’s tastes when they have never explicitly ranked two options?

  • BLUE learns to write user profiles that help predict what someone will choose next.
  • PUGC turns what people write into preference pairs for training language models.
  • PerRecBench tests whether a recommender can distinguish individual taste from item quality and rating habits.

A profile can also guide adaptation. BLUE asks what the profile should say. Profile-to-PEFT asks how to turn that description into adapter weights.

A model for each person

How much learning has to happen again when a new user arrives?

  • OPPU trains a small adapter for each user while keeping the base model frozen.
  • Per-Pcs lets users share and assemble adapters.
  • Profile-to-PEFT generates adapter weights directly from a new user’s profile.

Profile-to-PEFT and HyperHarness share a design choice. Both learn a generator across users, then run it once for a new person. One generates adapter weights, the other generates agent code.

Memory for individuals and teams

Which details from past interactions matter for the decision at hand, and whose perspective do they represent?

  • HyperHarness learns across users to build a retrieval and answer strategy tailored to each person’s history.
  • GroupBrief reconstructs a team’s working context, including who owns a decision and where people still disagree.

Foundations in user modeling. My earlier work studied individual accounts in TwiBot-22 and communities in BotPercent. LLM-UM surveys how language models can represent users from text and social connections.

🔥What's New
  • [2026.11] New papers on personalized and collaborative agents: HyperHarness and GroupBrief.
  • [2026.9] BLUE and GCAD were accepted to NeurIPS 2026!👏
  • [2026.8] SafeCoDe was accepted to EMNLP 2026 Main!👏
  • [2026.4] 5 papers were accepted to ACL 2026!👏 Main: Profile-to-PEFT, Findings: LoGo, GraphRL, Agentic Conv. Search, and MTMCS-Bench.
  • [2026.3] Received Tinker and Lambda Research Grants. Thank you Thinking Machines Lab and Lambda!
  • [2026.1] Passed my oral candidacy exam and now a PhD candidate!🥂
  • [2026.1] DRIFT was accepted to ICLR 2026!👏
  • [2025.05] APCIR was accepted to CIKM 2025!👏
  • [2025.05] 4 papers were accepted to ACL 2025!👏 Main: PUGC, StepCo, MANU. Findings: CodeTaxo.
  • [2025.02] Excited to join Microsoft Office of Applied Research as a part-time research intern this spring and work with Dr. Pei Zhou and Dr. Mengting Wan!
  • [2025.01] PerRecBench is live on arXiv! We remove the user rating bias and item quality in user rating prediction and specifically focus on evlauating personalization. Come check out our work!
  • [2025.01] 2 papers were accepted to NAACL 2025 Main!👏 MLLMU-Bench, IHEval. See you in New Mexico!
  • [2024.09] 4 papers were accepted to EMNLP 2024!👏 Main: OPPU, Per-Pcs, ProCo. Findings: NLGift. See you in Miami!🏖️
  • [2024.07] Chain-of-Layer was accepted to CIKM 2024!👏
  • [2024.06] Per-Pcs is live on arxiv. Come check out our framework of personalizing LLM with collaborative efforts!
  • [2024.05] 4 papers were accepted to ACL 2024👏! Main: BotSay. Findings: SKU, DELL, and K-Crosswords.
  • [2024.02] OPPU is live on arxiv. Welcome to check out our work!
  • [2024.01] KGQuiz was accepted to WWW 2024. Huge congrats to Yuyang!👏
  • [2024.01] I will join Amazon as an applied scientist intern this summer. See you in Palo Alto!
  • [2023.12] LLM-UM was accepted to DEBULL and live on arxiv. Welcome to check out our work and the reading list!🤗
  • [2023.10] LMBot was accepted to WSDM 2024. Huge congrats to Zijian!👏
  • [2023.10] BotPercent and MVSD were accepted to EMNLP 2023. Congrats to co-authors!
  • [2023.09] NLGraph was accepted to NeurIPS 2023 as a spotlight. Huge congrats to Heng!👏
  • [2023.07] I graduated from XJTU and stepped down as the LUD lab director. Be the Light of the World!🎓
  • [2023.05] KALM was accepted to ACL 2023, kudos to coauthors!👏
  • [2023.04] MVSD is live on arxiv, welcome to check out our work!
  • [2023.04] BotMoE was accepted to SIGIR 2023. Huge congrats to Yuhan!
  • [2023.03] I will join the University of Notre Dame to work with Prof. Meng Jiang this fall. Thank you for seeing my potential and looking forward to the incoming PhD journey!🥳
  • [2023.02] BotPercent is alive on arxiv, welcome to check out our work!
  • [2023.01] KRACL was accepted by WWW 2023, cheers!🍻
  • [2022.11] Our team CogDL-kgTransformer won the 4-th place in the OGB-LSC@NeurIPS2022 competition WikiKG90Mv2 track!
Selected Publications (* indicates equal contribution) [Google Scholar]
2026
Comparison of per-user harness optimization and direct generation with HyperHarness HyperHarness: Learning to Generate Personalized Agent Harnesses from User Histories
Zhaoxuan Tan, Jiachen Zhao, Jingbo Yang, Zheyuan Liu, Chuhan Wang, Pei Zhou, Mengting Wan, Meng Jiang

HyperHarness learns to generate retrieval and answer-guidance code from each user’s history. A shared generator builds a harness for a new user in one pass, without running a separate optimization loop for that user.

An example of a group briefing for an agent joining an ongoing project Grounding Collaborative Agents in Group Cognition
Zhaoxuan Tan, Mengting Wan, Jingbo Yang, Kwei-Herng Lai, Renyi Qu, Lorenzo Jaime Flores, Meng Jiang, Longqi Yang, Victor Poznanski, Pei Zhou

We introduce group briefing, a task that asks an agent to reconstruct what team members believe, who owns which decisions, and where shared understanding breaks down. We evaluate GroupBrief on 1,351 requests across four domains.

3DSP Bridging Textual Profiles and Latent User Embeddings for Personalization
Zhaoxuan Tan, Xiang Zhai, Yan Zhu, Meng Jiang, Mohamed Hammad
Proceedings of NeurIPS 2026.

We proposed BLUE, a reinforcement learning framework that aligns LLM-generated textual user profiles with embedding-based recommendation objectives, enabling interpretable and transferable user representations for zero-shot recommendation and personalized question answering.

3DSP Instant Personalized Large Language Model Adaptation via Hypernetwork
Zhaoxuan Tan, Zixuan Zhang, Haoyang Wen, Zheng Li, Rongzhi Zhang, Pei Chen, Fengran Mo, Zheyuan Liu, Qingkai Zeng, Qingyu Yin, Meng Jiang
Proceedings of ACL 2026.
project page

We introduce Profile-to-PEFT, a scalable framework that employs a hypernetwork to map user profiles directly to PEFT adapter parameters, enabling instant, efficient LLM personalization without per-user training and outperforming traditional methods while using significantly fewer computational resources.

2025
3DSP Aligning Large Language Models with Implicit Preferences from User-Generated Content
Zhaoxuan Tan, Zheng Li, Tianyi Liu, Haodong Wang, Hyokun Yun, Ming Zeng, Pei Chen, Zhihan Zhang, Yifan Gao, Ruijie Wang, Priyanka Nigam, Bing Yin, Meng Jiang
Proceedings of ACL 2025.
project page

We proposed PUGC, a framework leveraging implicit human preferences in unlabeled user-generated content (UGC) for scalable and domain-specific alignment of LLMs, achieving significant performance improvements over traditional methods on alignment benchmarks.

3DSP Can Large Language Models Understand Preferences in Personalized Recommendation?
Zhaoxuan Tan, Zinan Zeng, Qingkai Zeng, Zhenyu Wu, Zheyuan Liu, Fengran Mo, Meng Jiang
arXiv preprint 2025.

We proposed PerRecBench, a benchmark that isolates user rating bias and item quality to evaluate recommendation techniques in a grouped ranking manner, revealing that while larger LLMs perform better overall, they struggle with personalized recommendation, emphasizing the need for improved fine-tuning strategies and understanding of user preferences.

2024
3DSP Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts
Zhaoxuan Tan, Zheyuan Liu, Meng Jiang
Proceedings of EMNLP 2024.

We proposed Personalized Pieces (Per-Pcs) for personalizing large language models, where users can safely share and assemble personalized PEFT modules efficiently through collaborative efforts. Per-Pcs outperforms non-personalized and PEFT retrieval baselines, offering performance comparable to OPPU with significantly lower resource use, promoting safe sharing and making LLM personalization more efficient, effective, and widely accessible.

3DSP Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning
Zhaoxuan Tan, Qingkai Zeng, Yijun Tian, Zheyuan Liu, Bing Yin, Meng Jiang
Proceedings of EMNLP 2024.

We proposed One PEFT Per User (OPPU) for personalizing large language models, where each user is equipped a personal PEFT module that can be plugged in base LLM to obtain their personal LLM. OPPU exhibits model ownership and enhanced generalization in capturing user behavior patterns compared to existing prompt-based LLM personalization methods.

3DSP LMBot: Distilling Graph Knowledge into Language Model for Graph-less Deployment in Twitter Bot Detection
Zijian Cai, Zhaoxuan Tan, Zhenyu Lei, Zifeng Zhu, Hongrui Wang, Qinghua Zheng, Minnan Luo
Proceedings of WSDM, 2024.

We propose LMBot, which utilizes a language model with graph-aware knowledge distillation to act as a proxy for graph-less Twitter bot detection inference. This approach effectively resolves graph data dependency and sampling bias issues.

2023
3DSP User Modeling in the Era of Large Language Models: Current Research and Future Directions
Zhaoxuan Tan, Meng Jiang
IEEE Data Engineering Bulletin (DEBULL), 2023.
reading list

We summarize existing research about how and why LLMs are great tools of modeling and understanding UGC. Then we review a few categories of large language models for user modeling (LLM-UM) approaches that integrate the LLMs with text and graph-based methods in different ways. Then we introduce specific LLM-UM techniques for a variety of UM applications. Finally, we present remaining challenges and future directions in the LLM-UM research.

3DSP BotPercent: Estimating Bot Populations in Twitter Communities
Zhaoxuan Tan*, Shangbin Feng*, Melanie Sclar, Herun Wan, Minnan Luo, Yejin Choi, Yulia Tsvetkov
Proceedings of EMNLP-Findings, 2023.
demo / tweet

We introduce the concept of community-level Twitter bot detection and develope BotPercent, a multi-dataset, multi-model Twitter bot detection pipeline. Utilizing BotPercent, we investigate the presence of bots in various Twitter communities and discovered that bot distribution is heterogeneous in both space and time.

3DSP Can Language Models Solve Graph Problems in Natural Language?
Heng Wang*, Shangbin Feng*, Tianxing He, Zhaoxuan Tan, Xiaochuang Han, Yulia Tsvetkov
Proceedings of NeurIPS 2023 (Spotlight)
code

Are language models graph reasoners? We propose the NLGraph benchmark, a test bed for graph-based reasoning designed for language models in natural language. We find that LLMs are preliminary graph thinkers while the most advanced graph reasoning tasks remain an open research question.

3DSP HOFA: Twitter Bot Detection with Homophily-Oriented Augmentation and Frequency Adaptive Attention
Sen Ye, Zhaoxuan Tan, Zhenyu Lei, Ruijie He, Hongrui Wang, Qinghua Zheng, Minnan Luo
arXiv preprint 2023.

We identify the heterophilous disguise challenge in Twitter bot detection and proposed HOFA, a novel framework equipped with Homophily-Oriented Augmentation and Frequency Adaptive Attention to demystify the heterophilous disguise challenge.

3DSP BotMoE: Twitter Bot Detection with Community-Aware Mixtures of Modal-Specific Experts
Yuhan Liu, Zhaoxuan Tan, Heng Wang, Shangbin Feng, Qinghua Zheng, Minnan Luo
Proceedings of SIGIR 2023.

We propose community-aware mixture-of-experts to address two challenges in detecting advanced Twitter bots: manipulated features and diverse communities.

3DSP KRACL: Contrastive Learning with Graph Context Modeling for Sparse Knowledge Graph Completion
Zhaoxuan Tan, Zilong Chen, Shangbin Feng, Qingyue Zhang, Qinghua Zheng, Jundong Li, Minnan Luo
Proceedings of The Web Conference (WWW), 2023.
code / talk

We adopt contrastive learning and knowledge relational attention network to alleviate the widespread sparsity problem in knowledge graphs.

2022
3DSP TwiBot-22: Towards Graph-Based Twitter Bot Detection
Shangbin Feng*, Zhaoxuan Tan*, Herun Wan*, Ningnan Wang*, Zilong Chen*, Binchi Zhang*, Qinghua Zheng, Wenqian Zhang, Zhenyu Lei, Shujie Yang, Xinshun Feng, Qingyue Zhang, Hongrui Wang, Yuhan Liu, Yuyang Bai, Heng Wang, Zijian Cai, Yanbo Wang, Lijing Zheng, Zihan Ma, Jundong Li, Minnan Luo
Proceedings of NeurIPS, Datasets and Benchmarks Track, 2022.
website / GitHub / bibtex / poster

We present TwiBot-22, the largest graph-based Twitter bot detection benchmark to date, which provides diversified entities and relations in Twittersphere and has considerably better annotation quality.

3DSP Heterogeneity-Aware Twitter Bot Detection with Relational Graph Transformers
Shangbin Feng, Zhaoxuan Tan, Rui Li, Minnan Luo
Proceedings of AAAI 2022.
slides / code / bibtex

We propose the relational graph transformers GNN architecture to leverage the intrinsic relation heterogeneity and influence heterogeneity in Twitter network.

Industrial Experience
Microsoft
2026.05 - present

Research Intern @ Office of Applied Research
Host: Dr. Pei Zhou, Dr. Mengting Wan
Redmond, WA
Google
2025.10 - 2026.04

Student Researcher @ Google CoreML
Host: Dr. Yan Zhu, Dr. Xiang Zhai
Remote
Amazon Science
2025.06 - 2025.10

Applied Scientist Intern @ Rufus
Host: Dr. Zixuan Zhang, Dr. Haoyang Wen
Palo Alto, CA
Microsoft
2025.03 - 2025.06

Research Intern @ Office of Applied Research
Host: Dr. Pei Zhou, Dr. Mengting Wan
Remote
Amazon Science
2024.05 - 2024.10

Applied Scientist Intern @ Rufus
Host: Dr. Zheng Li, Dr. Tianyi Liu
Palo Alto, CA
Education
University of Notre Dame
2023.08 - present

Ph.D. in Computer Science and Engineering
Advisor: Prof. Meng Jiang
Xi'an Jiaotong University
2019.08 - 2023.07

B.E. in Computer Science and Technology
Advisor: Prof. Minnan Luo
Service
  • Area Chair/SPC: ARR (Feb-May 2025), AAAI 2027, WWW 2027
  • Reviewer: TMLR (2025), ICML (2025), AISTATS (2025), COLM (2024, 2025), KDD (2024, 2025), ARR (Dec 2023-), WWW (2024, 2025), ICLR (2024, 2025), TKDE (2023-), TNNLS (2023-), ICWSM (2024), NeurIPS (2023-2025), NeurIPS D&B (2022), LoG (2022, 2023, 2024), SRW@ACL (2025), AGI@ICLR (2024), TGL@NeurIPS (2023), GCLR@AAAI (2024), KnowledgeNLP@ACL (2024), WiNLP@EMNLP (2024).
  • Volunteer: EMNLP 2023 (virtual), EMNLP 2022 (virtual).
  • Director of the LUD lab (promote undergraduate research) 2022-2023.
Miscellaneous
  • I have the fortune to work with brilliant mentors, collaborators, and advisors during my research journey and I am truly grateful for their guidance and help. If you feel like I can be of some help to your research career, welcome to reach out!☕
  • I enjoy playing trumpet🎺 and served as the principal trumpet player in my primary and high school band🎼.
  • I am a big fan of Jules Verne, and especially fascinated with In Search of the Castaways, From the Earth to the Moon, and Five Weeks in a Balloon.
  • I also love jogging🏃 and playing table tennis🏓.
  • My favorite singers are Khalil Fong🕯️ and Jonathan Lee.

Template courtesy: Jon Barron.