Office:5544 Leinweber, 2200 Hayward St., Ann Arbor, MI 48109
Phone: 734-764-5876
Email: lastname followed by the letter 'p' at umich dot edu
Twitter: @dhillon_p
Google Scholar: https://goo.gl/FEsnE8
Quick Navigation Links
Please follow the links below to navigate to specific subsections of the site or just scroll down to view all the content.
My current research focuses on how Large Language Models (LLMs) reshape human creativity, decision-making, and information consumption. My work bridges machine learning, causal inference, and human-computer interaction (HCI) to understand and shape the societal impacts of language technologies.
Key Research Areas:
Human-LLM Collaboration/Co-writing and Creative Labor Markets: We investigate how people create with LLMs, examining questions of authorship, ownership, and market disruption [CHI '24 + under review]. Our empirical work provides the first systematic evidence that fine-tuned LLMs can produce undetectable, professional-quality writing that competes directly with human authors.
Personalization and Human-Centric Recommender Systems: We design LLM-based systems that respect human agency while providing personalized experiences. This includes developing temptation-aware recommendation algorithms that help users navigate between immediate desires and long-term goals, and creating personalization methods that adapt to individual preferences without reinforcing filter bubbles [WWW '24]. Our work on multi-stage policy learning enables LLMs to optimize sequential text-based interventions in domains like mental health support and educational content delivery.
Causal Methods for Language-Based Interventions: We advance causal inference techniques for high-dimensional text treatments [NAACL '24]. Our recent work introduces policy learning frameworks for natural language action spaces, enabling LLMs to learn optimal intervention strategies through gradient-based optimization on language embeddings. This methodology supports applications from therapeutic dialogue refinement to content moderation, where each text-based decision impacts future outcomes.
Active Research Threads:
Empirical measurement of LLM impacts on creative labor markets and copyright
Temptation-aware recommender systems that balance engagement with user well-being
Multi-stage decision-making with natural language actions
Psychological ownership and agency in human-LLM collaboration
Causal frameworks for evaluating text-based interventions
I received my Ph.D. in Computer Science from the University of Pennsylvania (2015), where I was advised by Professors Lyle Ungar, Dean Foster, and James Gee. My dissertation, "Advances in Spectral Learning with Applications to Text Analysis and Brain Imaging," received the Morris and Dorothy Rubinoff Award for outstanding doctoral dissertation. This work introduced theoretically-grounded spectral methods for learning word embeddings (JMLR 2015, ICML 2012, NeurIPS 2011) and brain image segmentation (NeuroImage 2014), achieving both computational efficiency and provable convergence guarantees. My contributions to spectral decomposition and context-dependent representation learning provided early theoretical foundations for understanding how distributed representations capture semantic relationships-- principles that remain central to modern transformer architectures. I also did other research in my Ph.D. on establishing connections between PCA and ridge regression (JMLR 2013) and on provably faster row and column subsampling algorithms for least squares regression (NeurIPS 2013a,b). I also hold an A.M. in Statistics and M.S.E. in Computer Science from Penn.
Following my Ph.D., I completed a postdoctoral fellowship at MIT with Professor Sinan Aral, where I worked on problems at the intersection of Machine Learning, Causal Inference, Network Science, and Information Systems. This research program produced several foundational contributions: establishing tractable methods for influence maximization under empirically-grounded network models (Nature Human Behaviour 2018); designing optimal digital paywall strategies that balance subscription revenue with content demand leveraging quasi-experiments (Management Science 2020); developing neural matrix factorization techniques for modeling temporal dynamics in user preferences (Marketing Science 2021); quantifying the information advantages of network brokers through novel diversity metrics (Management Science 2023); and creating surrogate-index methods for optimizing long-term outcomes in sequential decision problems (Management Science 2024). This body of work established rigorous approaches for causal inference in high-dimensional, networked settings.
ECE 4760J Data Mining: Methods and Applications @ UM-Shanghai Jiao Tong University Joint Institute (JI) @ S25.
SI 671/721 Data Mining: Methods and Applications @ F[19-24].
SIADS 642 [online] Deep Learning I [Developed from scratch] @ F20-present.
SIADS 647 [online] Deep Learning II (Generative AI) [Developed from scratch] @ W25.
SIADS 532,632 [online] Data Mining I, II @ W21-present.
Research Group
Ph.D. Students
Yachuan Liu [F20-] Last Stop: BS @ UC Berkeley.
Sanzeed Anwar [F21-] Last Stop: BS+MEng @ MIT.
Bohan Zhang [F22-] Last Stop: MS @ University of Michigan.
Siqi Liang [W25-] Last Stop: MS @ University of Southern California.
Former Students
Xinyue Li[MS '23, next Ph.D. in Statistics at Boston University]
Ella Li[MS '23, next Ph.D. in CS at Northeastern University]
Siqi Ma [BS '23, next MS in Statistics at Stanford University]
Shaochun Zheng [BS '23, next MS in CS at UC San Diego]
Houming Chen [BS '23, next Ph.D. in CS at University of Michigan]
Yushi She [BS '23, next MS in CS at Georgia Tech]
Ted Yuan [BS '23, next MS in ECE at Carnegie Mellon University]
Evan Weissburg [BS '23, next Software Engineer at Jane Street Capital]
Arya Kumar [BS '23, next Software Engineer at Jane Street Capital]
Jupiter Zhu [BS '22, next MS in CS at Stanford University]
Tianyi Li [BS '22, next MS in INI at Carnegie Mellon University]
Xianglong Li [BS '22, next MS in CS at Yale University]
Florence Wu [BS '22, next MS in CS at Harvard University]
Yingzhuo Yu [BS '22, next MS in CS at UIUC]
Xingjian Zhang [BS '22, next Ph.D. in Information at UMSI]
Bohan Zhang [MS '22, next Ph.D. in Information at UMSI]
Zhengyang Shan [MS '22, next Ph.D. in CDS at Boston University]
Jiapeng Guo [BS '21, next MS in CS at Columbia University]
Zilu Wang [BS '21, next MS in MS&E at Stanford University]
I always have openings for strong students in my group at all levels (Postdoctoral, Ph.D, Masters, or Undergrad). I am broadly looking to supervise students who are interested in working on Human-centric AI, Information Systems, or NLP. Prior research experience in these areas is highly valued, as are strong programming skills and a solid applied math/statistics background.
Process: Masters/Undergrads (already at University of Michigan) interested in working with me can email their CV and transcripts. Prospective Postdocs can directly email me their latest CV and Research Statement. Prospective Ph.D. students need not email me directly but are encouraged to apply to our Ph.D. program here and mention my name as a potential advisor. The deadline is December 1 each year.
Awards
INFORMS Information Systems Society (ISS) Gordon B. Davis Young Scholar Award, 2021.
INFORMS Annual Conference (Best Paper Award), 2020.
Workshop on Information Systems and Economics (WISE) (Runner-up Best Paper Award), 2016.
Rubinoff Best Doctoral Dissertation Award (awarded by Penn CIS), 2015.
Ad-hoc Reviewer: Nature, Nature Human Behaviour, Nature Communications, PNAS, JAIR, Information Science Research (ISR), Management Science, Marketing Science, IEEE TKDE, IEEE TPAMI.
Reviewer/PC/SPC Member @ Core AI/ML Conferences: [every year since 2013] NeurIPS, ICML, AISTATS, ICLR, AAAI, IJCAI.
Reviewer/PC/SPC Member @ Core Information Systems Conferences: [every year since 2017] ICIS, CIST, WISE.
Reviewer/PC/SPC Member @ Core NLP/Computational Social Science Conferences: [sporadically] EMNLP, NAACL, ICWSM, IC2S2.
Selected Publications
Below is a list of selected publications that highlight my core research interests and contributions. A complete list of all my publications is available here.
*indicates alphabetical author listing.
How Digital Paywalls Shape News Coverage.
Paramveer Dhillon, Anmol Panda, and Libby Hemphill.
PNAS Nexus, 2025.
[PDF]
Causal Inference for Human-Language Model Collaboration.
Bohan Zhang, Yixin Wang, and Paramveer Dhillon.
NAACL(Main Conference) (Annual Conference of the North American Chapter of ACL), 2024.
[PDF]
Shaping Human-AI Collaboration: Varied Scaffolding Levels in Co-writing with Language Models.
Paramveer Dhillon, Somayeh Molaei, Jiaqi Li, Maximilian Golub, Shaochun Zheng, and Lionel Robert.
CHI (SIGCHI Conference on Human Factors in Computing Systems), 2024.
[PDF]
Filter Bubble or Homogenization? Disentangling the Long-Term Effects of Recommendations on User Consumption Patterns.
Sanzeed Anwar, Grant Schoenebeck, and Paramveer Dhillon.
WWW (The Web Conference), 2024.
[PDF]
Targeting for long-term outcomes.
Jeremy Yang, Dean Eckles, Paramveer Dhillon, and Sinan Aral.
Management Science, 2023.
[PDF]
What (Exactly) is Novelty in Networks? Unpacking the Vision Advantages of Brokers, Bridges, and Weak Ties.
Sinan Aral, Paramveer Dhillon.
Management Science, 2022.
[PDF]
Modeling Dynamic User Interests: A Neural Matrix Factorization Approach.
Paramveer Dhillon, Sinan Aral.
Marketing Science, 2021.
[PDF]
Digital Paywall Design: Implications for Content Demand & Subscriptions.*
Sinan Aral, Paramveer Dhillon.
Management Science, 2020.
[PDF]
Social Influence Maximization under Empirical Influence Models.*
Sinan Aral, Paramveer Dhillon.
Nature Human Behaviour, 2018.
[PDF][Supplementary Information]
Eigenwords: Spectral Word Embeddings.
Paramveer Dhillon, Dean Foster, and Lyle Ungar.
JMLR (Journal of Machine Learning Research), 2015.
[PDF]
New Subsampling Algorithms for Fast Least Squares Regression.
Paramveer Dhillon, Yichao Lu, Dean Foster, and Lyle Ungar.
NeurIPS (Advances in Neural Information Processing Systems Conference), 2013.
[PDF][Supplementary Information]
Faster Ridge Regression via the Subsampled Randomized Hadamard Transform.
Yichao Lu, Paramveer Dhillon, Dean Foster, and Lyle Ungar.
NeurIPS (Advances in Neural Information Processing Systems Conference), 2013.
[PDF][Supplementary Information]
A Risk Comparison of Ordinary Least Squares vs Ridge Regression.
Paramveer Dhillon, Dean Foster, Sham Kakade, and Lyle Ungar.
JMLR (Journal of Machine Learning Research), 2013.
[PDF]
Two Step CCA: A new spectral method for estimating vector models of words.
Paramveer Dhillon, Jordan Rodu, Dean Foster, and Lyle Ungar.
ICML(International Conference on Machine Learning), 2012.
[PDF][Supplementary Information]
Multi-View Learning of Word Embeddings via CCA.
Paramveer Dhillon, Dean Foster, and Lyle Ungar.
NeurIPS(Advances in Neural Information Processing Systems Conference), 2011.
[PDF][Supplementary Information]
Minimum Description Length Penalization for Group and Multi-Task Sparse Learning.
Paramveer Dhillon, Dean Foster, and Lyle Ungar.
JMLR (Journal of Machine Learning Research), February 2011.
[PDF]
Software
Code and data for our Nature Human Behaviour 2018 paper is available here.
The ANTsR toolkit for medical image analysis (including the implementation of our NeuroImage 2014 paper) is available here.
The SWELL (Spectral Word Embedding Learning for Language) JAVA toolkit for inducing word embeddings (cf. JMLR 2015, ICML 2012, NeurIPS 2011) is available here.
Various Eigenword (SWELL) embeddings for reproducing the results in our JMLR 2015 paper can be found below [No additional scaling required for embeddings. Use them as is]. [Based on our results, OSCCA and TSCCA embeddings are the most robust and work best on a variety of tasks.]
OSCCA (h=2) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=2]
TSCCA (h=2) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=2]
LR-MVL(I) (h=2) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=2]
LR-MVL(II) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, smooths=0.5
OSCCA (h=10) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=10]
TSCCA (h=10) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=10]
LR-MVL(I) (h=10) [Trained on Reuters RCV1 (No lowercasing or
cleaning). v=100k, k=200, context size (h)=10]
Generic eigenwords embeddings for various languages [Trained on much larger corpora.]
English [Trained on English Gigaword (No lowercasing or
cleaning). v=300k, k=200]
German [Trained on German Newswire (No lowercasing or
cleaning). v=300k, k=200]
French [Trained on French Gigaword (No lowercasing or
cleaning). v=300k, k=200]
Spanish [Trained on Spanish Gigaword (No lowercasing or
cleaning). v=300k, k=200]
Italian [Trained on Italian Newswire+Wiki (No lowercasing or
cleaning). v=300k, k=200]
Dutch [Trained on Dutch Newswire+Wiki (No lowercasing or
cleaning). v=300k, k=200]