PhD candidate in Computer Science at the University of Iowa
I work with Muchao Ye on memory systems for large language models. A model can only reason over what fits in its context window, so memory systems save past interactions outside the model and bring back the relevant pieces when they're needed. I focus on making those systems efficient, accurate, and hard to manipulate, and on using them beyond chat, in settings like long videos.
Right now I'm working on training memory write policies with reinforcement learning, on using human feedback so a system's memory adapts to the person using it, on test-time learning, and on interfaces that show people what a system remembers so they can check it and trust it.
Alongside my PhD, I'm an AI development intern at Inspire Medical Systems, where I build LLM tools used by more than a dozen teams across the company. Before Iowa, I studied computer science at the University of St. Thomas in St. Paul and captained the Division I track and field team.
Efficiency and accuracy
A long-term memory system that saves each conversation turn word for word without running it through an LLM, then does the careful work at retrieval time. It had the best overall results on LoCoMo, LongMemEval-S, and MemGallery.
3.9 to 52.6× faster memory construction than the systems we compared against
Safety
Anyone who can chat with a memory-augmented LLM can write to its memory. ER-MIA plants false memories built to be retrieved right alongside real ones, then measures how much they change the model's answers.
17.7 to 28.5 points of accuracy lost on A-MEM across three models, from 3B to 27B parameters
Video
Gives a frozen vision-language model a memory of what it has already seen in a video, checks that memory against the current frames before using it, and groups evidence into whole events with one explanation each.
80.5% of the time, people preferred its explanations to clip-by-clip ones
Under review, 2026
Under review, 2026
Under review, 2026
IEEE International Conference on Machine Learning and Applications (ICMLA), 2025
May 2026 to present
Inspire Medical Systems · Minneapolis, Minnesota
I built and deployed an internal web app for an LLM document processing tool, which cut processing time by about 80% for more than 12 teams. I also build and evaluate RAG pipelines and AI agents that automate work across the company, and I bring ideas from research into production systems.
Aug 2025 to present
University of Iowa · Iowa City, Iowa
Research on memory for LLMs and multimodal LLMs with Muchao Ye. I lead MemFit, ER-MIA, and LATERN, and I built the evaluation harnesses we use to benchmark memory and RAG systems with F1, BLEU, ROUGE, exact match, and LLM-as-judge scoring.
May to Aug 2024
University of Colorado Colorado Springs
Worked on math reasoning in LLMs. I designed EVoSS, which has the model write equations for a symbolic solver and then checks the result against the model's own estimate. It became a first-author paper at IEEE ICMLA 2025. I also built two new benchmark datasets, SVAMPClean and Trig300.
2025 to present
University of Iowa
PhD candidate advised by Muchao Ye, supported by a Sheets Fellowship. I'll also receive an M.S. in Computer Science along the way in spring 2027.
2025
University of St. Thomas · St. Paul, Minnesota
Summa cum laude, 3.98 GPA. Minors in data science, applied statistics, philosophy and science, and logic and analytical reasoning. Captain of the Division I track and field team, all-conference athlete, holder of seven school records, and a member of the Student-Athlete Advisory Committee.
Aug 2024 to May 2025
University of St. Thomas
Developed materials for a graduate Foundations of AI course, including the interactive history of AI listed below, and researched how thinking about AI changed from early symbolic systems to neural networks. I was also a TA for Applied Regression Analysis.
Aug 2023 to May 2024
University of St. Thomas
Tutored students in Java, Python, C++, and MATLAB and gave feedback on their programming projects.
mitchellpiehl.github.io
An interactive timeline of the people, papers, and ideas in the history of AI, built for a graduate Foundations of AI course.
tosustainableai.com
A public guide to the environmental and social costs of AI and the work being done to reduce them.