Speaker: Pinar Yanardag Delul, assistant professor at Virginia Tech
Date and Time: October 9, 2:00-3:00 p.m. 
Location: Coda Building, Room 114
Host: Polo Chau

Title: Rethinking Memory: From Long-Horizon Video to Interactive World Models

Abstract: Video generation has made remarkable progress in visual quality, yet turning short clips into continuously evolving, controllable experiences remains a fundamental challenge. A model must preserve what matters from the past while allowing new events to unfold, respond to user input, and remain computationally efficient as generation unfolds. In this talk, I will explore these challenges through a central question: What should a video model remember, and how should it use that memory?

I will present three complementary approaches to rethinking memory in autoregressive video generation. First, Infinity-RoPE (CVPR’26) revisits how temporal context is organized, enabling longer video generation and more responsive control through inference-time interventions. Next, AdaState (NeurIPS’26) replaces a frozen anchor frame with an evolving latent state, balancing temporal consistency with dynamic scene progression. Finally, VideoMLA (NeurIPS’26) rethinks how video models store and access past context, substantially compressing their memory while preserving generation quality.

I will close with broader directions toward controllable world models (SPAWN, NeurIPS’26), physically grounded (HiPhy, NeurIPS’26) and diverse (DPP-GRPO, CVPR’26) video generation models. This research advances a broader vision for generative cinematography, supported by an NSF CAREER’26 award, toward generative worlds that remember the past, respond to our choices, and evolve as we interact with them.

Bio: Pinar Yanardag is an Assistant Professor of Computer Science at Virginia Tech, where she leads GEMLAB (Generative Modeling Lab). Her research focuses on controllable and personalized generative AI, with an emphasis on video generation and interactive world models. Her goal is to build systems that give people greater creative control over generated content and experiences. Since joining Virginia Tech in Fall 2023, her lab has produced over 25 publications at venues including CVPR, ICCV, and NeurIPS. She completed her PhD at Purdue as a Fulbright Fellow and was a postdoctoral researcher at the MIT Media Lab, where her generative AI projects attracted over one million participants and coverage from TIME, The Washington Post, and CNN. She is a recipient of the NSF CAREER award for her research on generative cinematography, and her research has received support from industry partners including Google, Adobe, and Snap.