READ: What Makes a Pretrained VLA Teachable by Demonstration?
Show a pretrained VLA five demonstrations of a new task and it performs the task with no gradient steps, matching per-task fine-tuning on held-out tasks.
Minh Pham-Dinh Kahlert Fellow
PhD student making multimodal agents more capable and adaptable
Advised by Kenneth Marino · m.pham-dinh (at) utah.edu
“The authors claim this is the first work to address this problem. This is false (Schmidhuber et al, 1999).” — Reviewer 2 idea from my college buddy Jacob
01About
I am a PhD student at the University of Utah, advised by Professor Kenneth Marino. These days I am mostly interested in enhancing the general capabilities and adaptation of multimodal agents: agents that see, read, and act, and that can pick up new skills and adjust to unfamiliar environments without being retrained from scratch.
Before Utah, I completed my undergraduate degree at Colby College, where I conducted AI and assistive robotics research at the Davis Institute for Artificial Intelligence and the INSITE Lab. I’m always open to collaborations on agents, both on screens and in the physical world.
02News
- READ accepted to two CoRL 2026 workshops: Pretrain to Adapt and RoboMem.
- New preprint: MiniWeb, a skill-centric approach to generalist web agents. Now on arXiv.
- Top Reviewer at NeurIPS 2026 (top 8%). See you in Atlanta!
- New preprint: WebMix, remixing adapters instead of retraining.
- Honored to receive the Kahlert Fellowship.
[older news]
- Started my PhD at the University of Utah.
- Presented DAVIS at EMNLP 2025.
- DAVIS accepted to Findings of EMNLP 2025.
- Dean’s List, Spring 2025.
- Presented DAVIS at the Colby Liberal Arts Symposium.
03Selected Publications
Back to Basics: A Skill-Centric Approach to Generalist Web Agents
Compact browser agents succeed on unseen websites by practicing reusable interaction skills, then composing them with a planner to solve full tasks.
Don't Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks
Web agents adapt to unfamiliar websites and redesigns by scouting the interface and remixing existing adapters, with no retraining on the target.
DAVIS: Planning Agent with Knowledge Graph-Powered Inner Monologue
A science agent that plans by querying a structured, temporal knowledge graph in a multi-turn inner monologue, improving on 8 of 9 ScienceWorld subjects.
04Contact
- m.pham-dinh (at) utah.edu
- Office
- Kahlert School of Computing
Salt Lake City, UT 84115 - Phone
- 207-830-3901