Harsh
Kumar.
I am a PhD candidate in Computer Science at the University of Toronto, advised by Ashton Anderson. I study how everyday use of LLMs shapes people’s learning, creativity, and wellbeing. I use those findings to design better evaluations and post-training methods for AI systems.
I study what long-term repeated AI interactions leave behind in people.
Evaluations today largely capture what models can do, and increasingly what humans and AI accomplish together. My work asks about another harder-to-measure layer, regarding the lasting effects of these interactions on human capacities. As LLM use scales, even small effects on individual learning, creativity, and wellbeing may accumulate into broader social consequences. I study these questions through large-scale randomized experiments and other methods from social science.
My dissertation, Human Capacities in the Age of AI: Designing and Evaluating LLM Systems for Learning, Creativity, and Wellbeing, will be defended in Fall 2026. Along the way, I spent two summers with the Computational Social Science group at Microsoft Research NYC, working with Dan Goldstein, Jake Hofman, and David Rothschild on how math can be taught with LLMs, including one of the first large-scale RCTs on LLMs and education. I also spent a summer as an Applied Scientist Intern at Microsoft with Shamsi Iqbal, studying how LLMs can support organizational sensemaking and the future of work.
My work has been covered by Time, The Economist, The New York Times, Forbes, and Science News.
I am on the 2026–27 academic and industry job market, for roles in human–AI interaction and LLM evaluation/post-training. CV · Email
News
- Jul 2026 — I am attending ICML 2026 in Seoul to present our position paper, We Need Large Language Models Optimized For Our Well-Being. Say hi if you are around!
- Apr 2026 — Two papers at CHI 2026: When AI Gives Advice (on AI vs. human responses to wellbeing advice-seeking) and LLM Use and Critical Thinking under Time Constraints (led by Jiayin Zhi, with Mina Lee), the latter covered by Science News.
- Apr 2025 — Math Education with LLMs was accepted to AIED 2025.
- Mar 2025 — Human Creativity in the Age of LLMs received an Honorable Mention award at CHI 2025. I also discussed the work on the Teaching, Learning, and Everything Else podcast, and it was covered by TechXplore.
- Feb 2025 — LLM Agents for Improving Engagement with Behavior Change Interventions was accepted to CSCW 2025.
Research Themes & Representative Papers
Hundreds of millions of people now use AI in ways that are difficult to measure but consequential. My dissertation examines three of these domains. Read more about my research here · See the full list of papers here
Creativity & Cognition
LLM assistance can boost creative performance in the moment, but may reduce unassisted creative ability and the diversity of what people produce, even after they stop using LLMs.
Human Creativity in the Age of LLMs
CHI 2025 ● Honorable Mention ● PDF ● Code
Randomized experiments with 1,100 participants on divergent and convergent thinking, and what repeated LLM use does to unassisted creativity.
Human Thinking under Plural LLM Assistance
under review ● Code
Controlled experiments examining how systematically varied LLM personas affect mathematical problem-solving and open-ended writing. Inspired by mechanisms of social learning in classrooms.
LLM Use and Critical Thinking under Time Constraints
CHI 2026 ● Project Website
How the timing of LLM access affects critical thinking under time pressure.
Learning
Does help from an LLM build understanding that persists once the model is gone? I also design interventions that target outcomes beyond retention, such as learners’ self-confidence, motivation, and reliance on the support.
Math Education with LLMs: Peril or Promise?
AIED 2025
A randomized controlled trial of LLM tutoring with 1,800 learners investigating under what contexts LLMs can lead to better learning.
Guiding Students in Using LLMs in Supported Learning Environments
CSCW 2024
Two classroom field experiments investigating how to help students use LLMs for learning.
Supporting Self-Reflection at Scale with LLMs
Learning@Scale 2024
Randomized field experiments in classrooms on LLM-supported self-reflection.
Wellbeing & Mental Health
People increasingly bring personal struggles to LLMs.
When AI Gives Advice
CHI 2026 ● Code
Compares AI and human responses to real online advice-seeking, measuring perceived helpfulness, empathy, safety, and other support qualities.
LLM Agents for Improving Engagement with Behavior Change Interventions
CSCW 2025
Includes a four-week longitudinal study comparing reflective and informational agents for digital mindfulness.
Using Adaptive Bandit Experiments to Increase and Investigate Engagement in Mental Health
IAAI 2024
Field experiments with Mental Health America’s text-messaging program to personalize support content, reaching more than 10,000 young adults.
Position: We Need Large Language Models Optimized For Our Well-Being
ICML 2026 ● PDF
The position statement highlights my overall research agenda on well-being. We argue that we need a well-being mode for LLMs, and provide a framework to operationalize the tensions around well-being support (short-term vs. long-term outcomes, individual vs. collective interests, autonomy vs. guidance).
Diagnosing and Repairing Persona Collapse in LLM Advice
under review
Human experts shift among distinct advice-giving personas depending on the situation; frontier LLMs collapse over 90% of contexts into a single supportive default. We formalize this failure mode and repair it with inverse-process distillation.
Grants & Competitions
● DARPA AI Tools for Adult Learning — US$250,000 award for QuickTA, an LLM-based learning tool that has supported over 5000 students at the University of Toronto. I co-led the project with John Stamper (CMU) and Norman Bier (Open Learning Initiative)
● XPRIZE Digital Learning Challenge — US$500,000 grand prize; I led development of the contextual-bandit personalization system used for large-scale adaptive classroom experiments
I have also helped write grants with my mentors, including Florea.ai, a John Templeton Foundation–funded project on wellbeing (PIs: Louis Tay, Karina Vold, Ashton Anderson), where I lead the technical side as the senior graduate student.
Contact
I am always happy to chat! Email is the best way to reach me.
Website template adapted from Abhraneel Sarma.