Speaker: Vardan Papyan (University of Toronto)Title: Attention Sinks: A ‘Catch, Tag, Release’ Mechanism for EmbeddingsAbstract: Large language models (LLMs) often concentrate their attention on a small set of tokens—referred to as attention sinks. Common examples include the first token, a promptindependent sink, and punctuation tokens, which are prompt-dependent. Although these tokens often lack inherent semantic meaning, their presence is critical for model performance, particularly under model compression and KV-caching. Yet, the function, semantic role, and origin of attention sinks—especially those beyond the first token—remain poorly understood.In this talk, I’ll present a comprehensive investigation revealing that attention sinks catch a sequence of tokens, tag them with a shared perturbation, and release them back into the residual stream, where they are later retrieved based on the tags they carry. Probing experiments show that these tags encode semantically meaningful information, such as the truth of a statement.This mechanism persists in models with query-key normalization—where prompt-dependent, non-BOS sinks have become more common—and DeepSeek-distilled models, where it spans more heads and accounts for greater variance in the embeddings. To support future theoretical work, we introduce a minimal task that is solvable via the catch, tag, release mechanism, and in which the mechanism naturally emerges through training.Biography: Vardan Papyan is an Assistant Professor in the Department of Mathematics at the University of Toronto, cross-appointed with the Department of Computer Science. He completed his postdoctoral studies at the Department of Statistics at Stanford University, under the guidance of David Donoho, and his PhD at the Department of Computer Science at the Technion – Israel Institute of Technology, under the supervision of Michael Elad.

The Mobility Forum with Prof. Jinhua Zhao showcases transportation research and innovation across the globe. The Forum is online and open to the public.

The convergence of sensors, spatial interfaces, and large visual-language AI models are transforming how we perceive, understand, and act in the physical world. Unlike traditional computing paradigms, embodied systems share our viewpoint and real-time context - enabling seamless spatial interaction. In this talk, I present three themes from our research on the future of Physical AI—where human experience, spatial computing, and intelligent systems converge to augment physical understanding and action.First, I will discuss our work on authoring environments that empower non-programmers to easily create immersive extended-reality applications. I will provide examples from hands-on training, production, and education to show how this works. Our system, which we call agentAR, enables subject-matter experts to author spatial learning experiences using both voice and gesture that help novices create augmented reality (AR) applications at ease and scale. Second, I will highlight our advances in AI design generation that take in human input and generate designs in real time. Our platform, designfromX, integrates vision and language models to convert verbal prompts and sketches into 3D designs, allowing humans and AI to co-create 3D design models. Third, I will present applications of embodied Physical AI in task performance and skill augmentation. Our avaTTar system is an extended reality table tennis-playing coach that fuses immersive visual feedback, and coaching intelligence. Building on avaTTar I will present a table-tennis-playing humanoid, offering a glimpse into the future of embodied AI. I will end by discussing visuo-haptic interfaces and smart physical tools to train workers in hands-on manufacturing settings.Together, these systems point to a future where Physical AI enhances how we design, train, and learn—expanding human potential across engineering, production, sports, and beyond. By bridging immersive interfaces and embodied intelligence, we aim to shape a new class of accessible, real-time, and spatially aware engineering systems.Biography: Karthik Ramani is the Donald W. Feddersen Distinguished Professor of Mechanical Engineering at Purdue University, with additional appointments in Electrical and Computer Engineering and a courtesy role in the College of Education. He leads the Convergence Design Lab, where his research brings AI into the physical world by blending human-centered AI with spatial intelligence to create immersive, real-time solutions for design, manufacturing, sports training, surgery, and hands-on learning. His work spans augmented spatial interactions, symbiotic human-AI collaboration, computational design thinking and prototyping, and scalable upskilling platforms for production. Using the lens of Physical AI, he develops systems that perceive, understand, and act in real environments—extending human capacity through embodied and intuitive interfaces.He has recently published in top-tier venues across computer vision (CVPR, ECCV, ICCV), human-computer interaction (ACM CHI, UIST), and AI (NIPS, ICLR), in addition to leading engineering design journals. He co-founded VizSeek, the world’s first commercial shape-based search engine for mechanical parts, and ZeroUI, a CES-awarded robotics startup. His educational innovations include Purdue’s Toy Design and Product-Process-Business Model Design courses. He was a visiting professor at Stanford University and Oxford University and a research fellow at PARC (formerly Xerox PARC). He earned his B.Tech from IIT Madras, M.S. from The Ohio State University, and Ph.D. from Stanford—all in Mechanical Engineering. He currently also serves as coach of Purdue’s Table Tennis team, where research meets passion— in the emerging domain of Athletic AI.

Ten minutes for your mind@2:50 every day at 2:50 pm in multiple time zones:Europa@2:50, EET, Athens, Helsinki (UTC+2) (7:50 am EST) https://us02web.zoom.us/j/88298032734Atlantica@2:50, EST, New York, Toronto (UTC-4) https://us02web.zoom.us/j/85349851047Pacifica@2:50, PST, Los Angeles, Vancouver (UTC=7) (5:50 pm EST) https://us02web.zoom.us/j/85743543699Almost everything works better again if you unplug it for a bit, including your mind. Stop by and unplug. Get the benefits of mindfulness without the fuss.@2:50 meets at the same time every single day for ten minutes of quiet together.No pre-requisite, no registration needed.Visit the website to view all @2:50 time zones each day.at250.org or at250.mit.edu

Time: 5:00 PMLocation: Cambridge, MA

Bhagavad Gita Fall Lecture SeriesJoin HG Sadananda Dasa, MIT Vaishnava Hindu Chaplain, for a weekly journey into the timeless wisdom of the Bhagavad Gita. Explore profound questions about identity, purpose, karma, yoga, love, and selfless service, and discover practical insights for living a meaningful and spiritually grounded life. Each session combines reflection, discussion, and practical tools for self-realization.RSVP: tinyurl.com/mitgita25