I am a final year PhD candidate at University of Michigan, Ann Arbor. My research interests are at the intersection of Machine Learning and Human-Computer Interaction (HCI).

I develop adaptive AI systems that enable people to reason under risk and uncertainty in complex decision-making scenarios. For example, enabling decision-makers in education to conduct accurate assessments of students for admissions and course waivers or designing adaptive meeting recap that considers meeting participant's goals and expertise to provide them the most relevant information. I borrow from cognitive science to build approximate user models for tasks, and use them with Reinforcement Learning to improve Human-AI interaction. I bring in strong computational and model building skills from my prior industry experience to build systems for Human-AI interaction and my training in HCI allows me to conduct real world study on improving Human-AI interaction. E.g., I recently built a bayesian network from a massive dataset of 3M records to model personal information and using it to study personalization - privacy trade-off.

I am an open source contributor to Wikipedia, and I also administered their Google Summer of Code internship program in 2016,17. Specifically, I contributed majorly to the mobile Wikipedia website to make it high-performant. For my services I was nominated by Wikimedia Foundation to attend Google summer of code mentor summit in 2017, and invited to present my research at the monthly Wikimedia research showcase in 2021.

Sumit Asthana

asumit at umich dot edu

PhD candidate, Computer Science and Engineering
University of Michigan, Ann Arbor
Ann Arbor
United States

Google Scholar
Twitter
LinkedIn
CV

Publications

Evaluating LLMs for Targeted Concept Simplification for Domain-Specific Texts.

In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 6208–6226, Miami, Florida, USA. Association for Computational Linguistics.

Sumit Asthana, Hannah Rashkin, Elizabeth Clark, Fantine Huot, Mirella Lapata

Summaries, Highlights, and Action items - Design, implementation, and evaluation of an LLM-powered meeting recap system.

In Proceedings of ACM conferene on Computer Supported Cooperative Work (CSCW 2024)

Sumit Asthana, Sagi Hilleli, Pengcheng He, Aaron Halfaker

Understanding Uncertainty: How Lay Decision-makers Perceive and Interpret Uncertainty in Human-AI Decision Making

In Proceedings of ACM conference on intelligent user interfaces (IUI) 2023

Snehal Prabhudesai, Sumit Asthana, Leyao Yang, Xun Huan, Q. Vera Liao, Nikola Banovic

eneration and assessment of multiple-choice questions from video transcripts using large language models

In Proceedings of the Eleventh ACM Conference on Learning@ Scale (pp. 530-534).

Taimoor Arif, Sumit Asthana, Kevyn Collins Thompson

Automatically Labeling Low Quality Content on Wikipedia by Leveraging Patterns in Editing Behaviors

In Proceedings of the ACM on Human-Computer Interaction, Vol. 5, No. CSCW2, Article 359 (October 2021). ACM 23 pages

Sumit Asthana, Sabrina Tobar Thommel, Aaron Halfaker, Nikola Banovic

Whodo: Automating reviewer suggestions at scale

In Proceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. 2019. (FSE)

Sumit Asthana, Rahul Kumar, Ranjita Bhagwan, Christian Bird, Chetan Bansal, Chandra Maddila, Sonu Mehta, B. Ashok

Rex: Preventing bugs and misconfiguration in large services using correlated change analysis

In Proceedings of the 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI 20). 2020.

Sonu Mehta, Ranjita Bhagwan, Rahul Kumar, Chetan Bansal, Chandra Maddila, B. Ashok, Sumit Asthana, Christian Bird, Aditya Kumar

With Few Eyes, All Hoaxes are Deep

In Proceedings of the ACM on Human-Computer Interaction, Vol. 2, CSCW, Article 21 (November 2018). ACM, New York, NY. 18 pages. (Accpetance rate: 24%).

Sumit Asthana and Aaron Halfaker.

Industry Experience

Google Deepmind, New York

Research Intern, August 2023 - December 2023

Adaptive simplification of domain-specific writing using LLMs. Developed a robust semantic evaluation method to evaluate LLMs for simplifying domain-specific writing and evaluated open-source and commercial LLMs for their performance on the task.

Microsoft Research, Seattle

Research Intern, May 2022 - August 2022

Intelligent meeting recap - Designed, Built and Evaluated an LLM based meeting experiences and studied its effectivenss in the context of user's meeting. Insights informed to the development of a robust meeting recap exeperience within Microsoft.

Microsoft Research, Seattle

Research Intern, May 2021 - August 2021

Usefulness and challenges of devops bots in the software development landscape - Studied bots in a bot platform used by 10,000 developers daily for their workflows. Identified bot challenges such as too many recommendations, poorly grounded human-bot communication. Recommendation from the study helped improve bot engagement, and overall software development quality with Microsoft.

Microsoft Research, Bengaluru

Research Fellow, May 2018 - July 2019

Developed a reviewer recommendations platform within Microsoft. Evaluated the platform through a continual quantative user study, and interviewed developers to understand the breakdowns of the system, and potential direction for improvement such as understanding code level semantic knowledge of developers for review recommendation.

Wikimedia Foundation, India

Intern, Jan 2018 - June 2018

Built a system to identify high level topics of Wikipedia articles using word embeddings, and RandomForests. The models are currently deployed on Wikipedia, and helping Wikipedia editors identify articles of interest to judge relevance, and decide their value for Wikipedia.

Arista Networks, Bengaluru, India

Intern, July 2017 - May 2018

Worked on L2VPN routing solutions for Gigabit ethernet datacenter networks.

Projects

Designing adaptive assessments for understanding students knowledge in course learning

Personalized AI can improve education significantly by giving each student the opportunity to learn at their own pace and understanding. In this work, I am leveraging cognitive theory concept understanding to build an adaptive assessment methodology that assesses student competence while minimizing burden on them. I formulated a structured decision space of concepts necessary to understand the course, and combined LLM interaction with a submodular objective evaluation of concepts.

Alinging AI-assisted decision-making with end-user preferences

AI may make decisions that do not always align with end-user sensibilities. For example, during personalization, systems may collect data and infer and use sensitive attributes about their users without their awareness or consent. In this work, I engineered a Human-AI interactive system that collects the most optimal and informative data about its users for personalization using a submodular objective. In a between and within-subjects user study, I then demonstrated that users show differential preferences for the usage of sensitive attributes about them that the AI infers. Their preferences depend on the sensitivity and the accuracy of inferences. The study provides concrete evidence of incorporating explicit user feedback in systems for aligning AI objectives with user preferences.

Aligning Wikipedia article quality assessment by learning from policy-guided Wikipedia edits.

Content assessment is guided by content policies -- neutral point of view (NPOV), citations, and clear writing. For AI to understand Wikipedia content quality assessment, it needs to understand the application of the content policies in-context. I showed that Wikipedia edits that instantiate these content policies learn more aligned assessments of Wikipedia content policies than learning form sentences on highest quality articles.

Modeling indoor covid-19 transmissions. (Demo )

Built a system to evaluate the spread of covid-19 in indoor environments, by simulating human behavior using Reinforcement Learning. Through realistic models of home environments, and human behavior, we are understanding the spread of covid-19 due to human movements, to identify possible strategies to mitigate such spread. Click here for a demo video.

 

Plain Academic