Events

CILVR Seminar: Improving Agents via Online Interaction

Speaker: Nicklas Hansen

Location: 60 Fifth Avenue, Room 7th Floor Open Space
Videoconference link: https://nyu.zoom.us/j/99398421341

Date: Wednesday, February 4, 2026

Current learning systems are often trained on fixed datasets, yet real-world agents operate in a regime where the data they see depends on their policy. While this induces distribution shift, it also creates an opportunity: continual improvement through interaction. In this talk, I will connect a set of recent results showing that online interaction can be a powerful lever for iterative improvement of AI agents across world modeling, robotics, and language. Concretely, I will first cover our work on TD-MPC2 and Newt, highly data-driven approaches to world models that scale with data and model size, improve autonomously with online interaction across hundreds of tasks, and can be trained on a modest budget. Next, I will discuss how these ideas enable robots to acquire new skills directly in the real world through closed-loop data collection and adaptation. Finally, I will introduce recent work on post-training large language models with reinforcement learning from verifiable signals, including algorithmic insights that substantially improve the efficiency of GRPO-style methods. Throughout, I will highlight common principles, failure modes, and promising directions for building agents that learn via interaction.

 

Bio: Nicklas Hansen is a PhD candidate at University of California San Diego advised by Professors Xiaolong Wang and Hao Su. Their research focuses on developing generalist AI agents that learn from interaction. Nick has spent time at NVIDIA Research, Meta AI (FAIR), as well as Berkeley AI Research, and received their BS and MS degrees from Technical University of Denmark. They were a recipient of the 2024 NVIDIA Graduate Fellowship, and their work has been featured at top venues in machine learning and robotics. Webpage: www.nicklashansen.com