What should we measure?
I build task-grounded benchmarks and agentic quality graders for Copilot, including Today. EmailBench evaluates whether agents complete enterprise email and productivity tasks, beyond whether their tool calls succeed.
Human–agent interaction · Evaluation · Post-training
Hyeonsu B. Kang
Senior Applied Scientist, Microsoft
I focus on human–agent interaction, agent evaluation, and post-training for Copilot. I build task-grounded benchmarks, agentic quality graders, and synthetic training data for tool use, multimodal reasoning, and adversarial robustness—connecting model improvements to product quality, cost, and latency.
My research combines controlled experiments, behavioral measurement, and production evaluation to align agents with how people actually work. Previously, I developed learning tools at MIT and deployed scholarly discovery systems with AI2. With Conservation X Labs, I built team-formation algorithms for global innovation contests offering over $2M in prize funding.
I earned my Ph.D. in Human–Computer Interaction at Carnegie Mellon.
Research
How can evaluations improve agents?
I connect evaluation to synthetic data and post-training. VLM-SlideEval studies slide comprehension and perturbation sensitivity; my counterfactual red-teaming work turns everyday tasks into controlled adversarial cases.
What do people actually want?
I study how behavior can inform better objectives for AI. Users Mispredict examines where stated preferences diverge from choices; my Semantic Scholar work tests interventions with real users.
Publications
Google Scholar2026
Users Mispredict What Drives Their Preferences for AI Writing Assistance
* Equal contribution: Zana Buçinca and Hyeonsu B. Kang.
-
EmailBench: A Benchmark for Evaluating LLM Agents on Enterprise Email and Productivity Tasks
-
From Everyday Tasks to Red-Team Cases: Counterfactual Generation and Practical Defenses for AI Agents
-
PIKE: Graph-Enhanced Email Retrieval via Critic-Guided Expansion
2025
VLM-SlideEval: Evaluating VLMs on Structured Comprehension and Perturbation Sensitivity in PPT
BioSpark: Beyond Analogical Inspiration to LLM-augmented Transfer
Best Paper Honorable Mention
Inkspire: Supporting Design Exploration with Generative AI through Analogical Sketching
-
Hypoveil: A Privacy-Utility Trade-off Aware Framework for Collaborative Multi-Agent Reasoning
-
Investigating Explainability with Privacy-Preserving Multi-agent Systems
2024
Mitigating Barriers to Public Social Interaction with Meronymous Communication
Best Paper Award
PaperWeaver: Enriching Topical Paper Alerts by Contextualizing Recommended Papers with User-collected Papers
The Semantic Reader Project: Augmenting Scholarly Documents through AI-Powered Interactive Reading Interfaces
Imitation of Life: A Search Engine for Biologically Inspired Design
2023
2022
Augmenting Scientific Creativity with an Analogical Search Engine
Threddy: An Interactive System for Personalized Thread-based Exploration and Organization of Scientific Literature
From Who You Know to What You Read: Augmenting Scientific Recommendations with Implicit Social Networks
Scaling Creative Inspiration with Fine-Grained Functional Aspects of Ideas
Augmenting Scientific Creativity with Retrieval across Knowledge Domains
Earlier publications · 2017–2019
2019
Matching Open Innovation Projects for Analogical Feedback Exchange
2018
Paragon: An Online Gallery for Enhancing Design Feedback with Visual Examples
Custom Blocks in StarLogo Nova: A Template-Based Approach to Abstraction for Improved Ease of Use and Expressive Power