Harsh
Kumar.
I am a PhD candidate in Computer Science at the University of Toronto, advised by Ashton Anderson. I work on human–AI interaction and the evaluation and post-training of large language models, understanding how people use generative AI in everyday life, and how these systems can be measured and trained to have a better long-term impact.
I study what AI interaction leaves behind in people.
Evaluations today largely capture what models can do, and increasingly what humans and AI accomplish together. My work asks about a third, harder-to-measure layer: the lasting effects of these interactions on human capacities, in everyday uses, where the human stakes are high. For instance, how we learn, how we create, and how we seek well-being support. I study these questions through large-scale randomized experiments with real users, and explore evaluation and post-training methods informed by what the findings.
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. And a summer as an Applied Scientist Intern at Microsoft with Shamsi Iqbal, studying how LLMs can support organizational sensemaking, helping leaders make decisions and navigate AI adoption.
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. CV · Email
News
- May 2026 — Our position paper, We Need Large Language Models Optimized For Our Well-Being, was accepted to ICML 2026.
- Apr 2026 — Two papers at CHI 2026: When AI Gives Advice (on AI vs. human responses to well-being 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.
- Sep 2024 — Gave an invited talk at the Data Science for Mental Health SIG at The Alan Turing Institute.
- Aug 2024 — Transparency Disclosures and Reliability Disclaimers for Learner–LLM Interactions was accepted to HCOMP 2024.
- May 2024 — Guiding Students in Using LLMs was accepted to CSCW 2024, and Supporting Self-Reflection at Scale with LLMs to Learning@Scale 2024.
- Jan 2024 — Math Education with LLMs: Peril or Promise? was featured in Microsoft’s New Future of Work report.
Selected Papers
See the full list of papers here
Diagnosing and Repairing Persona Collapse in LLM Advice
request for preprint, under review
Position: We Need Large Language Models Optimized For Our Well-Being
ICML 2026 ● Link
Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking
CHI 2025 ● HONORABLE MENTION ● PDF
Covered by TechXplore and discussed on the Teaching, Learning, and Everything Else podcast.
When AI Gives Advice: Evaluating AI and Human Responses to Online Advice-Seeking for Well-Being
CHI 2026 ● Link
Guiding Students in Using LLMs in Supported Learning Environments: Effects on Interaction Dynamics, Learner Performance, Confidence, and Trust
CSCW 2024 ● PDF
Math Education with Large Language Models: Peril or Promise?
AIED 2025 ● PDF
One of the first large-scale randomized trials of LLM tutoring; featured in Microsoft’s New Future of Work report.
Selected Awards & Funding
● DARPA AI Tools for Adult Learning — US$250,000 award for QuickTA, an LLM-based learning tool that has supported 1,500+ students (2023)
● XPRIZE Digital Learning Challenge — US$500,000 grand prize, as part of the Adaptive Experimentation Accelerator team (2023)
Website template adapted from Abhraneel Sarma.