Industry Affiliates

Faculty Panel Discussion

Moderator: Magda Balazinska

Panelists: Taylor Kessler-Faulkner, Steve Seitz, Ben Shapiro, and James Weichert

Keynote Speaker: Stephanie Wang

Abstract: While AI agents have revolutionized software engineering, it remains very challenging to build state-of-the-art AI systems. This is in part because of the long tail of workloads in both training and inference, spanning many use cases, models, and hardware. Such diversity demands optimizations tailored to the workload. In this talk, I’ll discuss some of our ongoing work on harnessing coding agents to autonomously build reliable and production-quality AI systems.

Biography: Stephanie Wang joined the faculty of the Paul G. Allen School of Computer Science & Engineering at the University of Washington in fall 2024. Her research is primarily in distributed systems. She is especially interested in how we can build intermediate abstractions that make it easier to build high-performance, fault-tolerant, and domain-specific systems. Along those lines, she is interested in domains where large-scale application development is still pretty hard, such as machine learning and data processing. She’s also interested in programming languages as the interface that can improve distributed systems.

Previously, Wang was a Ph.D. student in the RISELab at UC Berkeley, where she was advised by Ion Stoica. She is also a co-creator and committer for the open-source project, Ray, which has been used to train ChatGPT, serve high-performance LLMs, and break the CloudSort 100TB record. She subsequently continued to develop the Ray ecosystem as a software engineer at Anyscale, working primarily on Ray Data, a system for distributed data preprocessing for ML.

Introduction and Overview, Moderator TBD

Symbolic Relations for Network Change Verification, Brody Franks ↗

Abstract: Relational network verification is a practical way to validate network changes without requiring exhaustive specifications, but existing tools compare a pair of concrete network states (pre- and post-change). They cannot answer what operators often care about: across the space of all possible network environments, which includes external route advertisements and failures, which behaviors could be affected by a change, and how?

To answer this question, we introduce Symbolic Relation Algebra (SRA), a new language and engine for network change verification. SRA can: (1) verify both the control plane and the data plane, in one general formalism, (2) customize the data representation at a low level of granularity, while retaining the same frontend specification, (3) support specialized data structures that hybridize concrete and symbolic verification and are efficient on real-world workloads.

SRA models network behaviors as binary relations whose fields may range over large domains. Input specifications are lowered to query plans over field-decomposed decision diagrams, propagating intermediate constraints lazily during specification checking. We demonstrate the effectiveness of our tool by applying SRA to representative network-change scenarios, examining both its expressiveness and performance. We show that SRA checks symbolic relational specifications 25-541× faster than a general-purpose solver.

Understanding Workloads and Optimizing KVCache Management for Coding Agents, Shubham Tiwari ↗

Abstract: Coding agents are a fast-growing LLM application, executing as long-running closed-loop sessions in which LLM generations alternate with external tool calls. Yet, unlike chat workloads, their serving behavior has not been studied extensively. We address this gap by collecting a dataset of real-world coding assistant traces. Our analysis shows that coding agent sessions repeatedly reuse large prefixes and create sustained KVCache pressure that conventional LLM serving policies handle poorly. Based on our analysis, we present CacheWise, a KVCache management layer that improves KVCache reuse for coding agent workloads. CacheWise combines prefix-aware scheduling with reuse-aware eviction guided by lightweight predictions from tool call metadata. We implement CacheWise on vLLM  and our evaluations on our collected traces show that it significantly reduces KVCache evictions and improves end-to-end session completion times.

Self-Defining Systems: Building and Operating Bespoke Systems with Agents, Shihang Li ↗

Abstract: Today’s systems software is built for generality, and the abstractions every user needs keep it from exploiting what any one deployment allows. Specialization has long been known to recover large gains, but building and maintaining a separate system for each deployment has rarely paid for itself. Coding agents are changing this. This talk presents two steps toward creating bespoke systems that agents build and operate for each deployment.

VibeSys is an agentic system that generates end-to-end systems. We show how to design a multi-agent system that designs, implements, and optimizes systems end to end. For example, VibeSys generates LLM serving systems from scratch for a given model, workload, and hardware target, outperforming production engines by up to 5.95× where general engines miss opportunities.

As bespoke systems are deployed, the Self-Defining Operator (SDO) improves systems operations. SDO is an agentic operator that operates distributed applications by learning from the incidents it resolves. For each incident, it writes a playbook and a detector, so later occurrences of the fault are caught and resolved cheaply. On recurring faults in Kubernetes applications, SDO cuts median time to resolution by 52% and token use by 59% compared with an agent that keeps no memory.

Introduction and Overview, Jennifer Mankoff

View the Grand Challenge that relates to this topic

ArtInsight: AI-Powered Artwork Engagement for Mixed Visual-Ability Families, Arnavi Chheda-Kothary ↗

Abstract: I will introduce ArtInsight, a novel AI-powered system to facilitate deeper engagement with child-created artwork in mixed visual-ability families. ArtInsight leverages large language models (LLMs) to craft a respectful and thorough initial description of a child’s artwork, and provides: creative AI-generated descriptions for a vivid overview, audio recording to capture the child’s own description of their artwork, and a set of AI-generated questions to facilitate discussion between blind or low-vision (BLV) family members and their children. We evaluated ArtInsight with five groups of BLV family members and their children. Our findings highlight a preference for ArtInsight’s longer, artistically-tailored descriptions over those generated by existing BLV AI tools. Participants highlighted the creative description and audio recording components as most beneficial, with the former helping “bring a picture to life” and the latter centering the child’s narrative to generate context-aware AI responses. I will also discuss the formative work that led to ArtInsight, as well as a deployment study where we observed real-world use cases of ArtInsight over several weeks in family and educational settings.

TaskAudit: Detecting Functiona11ity Errors in Mobile Apps via Agentic Task Execution, Mingyuan Zhong ↗

Abstract: Coming soon

Personal Fabrication for Soft Assistive Adaptations to Everyday Objects, Jerry Cao ↗

Abstract: Assistive adaptations often work best on objects people already own, but 3D printing onto existing objects usually needs precise scans, specialized printers, and leaves you with hard plastic. AugPrint prints soft, custom foam directly onto everyday objects using a consumer 3D printer and a smartphone scan. Plastic extruded from high above the surface coils into a compressible foam that drapes over curves and tolerates scan errors. Users can vary stiffness across a surface, add tactile textures, and embed pressure sensors. I’ll show a pressure-mapping bike seat and a gesture-sensing input pad, and discuss how this could make personalized assistive adaptations easier to create.

Introduction and Overview, Katharina Reinecke

View the Grand Challenge that relates to this topic

AI Safety Is a Multi-agent Problem, Carrie Yuan

Abstract: When we train AI, we assume the classic MDP paradigm, where the AI is training against a static environment. But the real world is not static. It’s full of humans who change and adapt as a result of using AI. As AI is deployed into the world, it alters both the people who use it (through social sycophancy and cognitive decline) and even our collective culture. One example is the artificial hivemind: language models increasingly produce the same homogeneous outputs. Another is the way LLMs distort our written language. In this talk, I will argue that AI safety is a multi-agent problem, and I will present MODA, a multi-agent approach to diverse generation. MODA treats each mode as a lightweight agent within a single shared policy, and it trains these agents to produce responses that are both high quality and diverse. This addresses a critical issue for building transparent and globally beneficial AI.

Open AI Infrastructure to Accelerate Science, Jacob Morrison ↗

Abstract: Scientists need to understand and control the tools they work with. Most capable AI systems are closed, so researchers can use them but cannot see how they were built, adapt them to their own data, or measure what they cost. At Ai2, we address this by releasing the full model flow for our Olmo models, including the training data, code, recipes, intermediate checkpoints, and evaluations for every stage of training. In this talk, I will argue that AI built for science needs this kind of openness, and I will describe three recent projects that build on it. Olmo Hybrid is our first hybrid model that we find trains more efficiently than a standard transformer. A study of Olmo 3’s development accounts for the energy, water, and compute behind every run, including all research and development. Branch-Adapt-Route adds new domain expertise to a trained model without retraining it from scratch. Together, these projects show how open models let researchers study, adapt, and specialize AI for their own needs.

African Data Ethics: A Discursive Framework for Black Decolonial AI Ethics, Teanna Barrett ↗

Abstract: The shift towards pluralism in global data ethics acknowledges the importance of including perspectives from the Global Majority to develop responsible data science practices that mitigate systemic harms in the current data science ecosystem. Sub-Saharan African (SSA) practitioners, in particular, are disseminating progressive data ethics principles and best practices for identifying and navigating anti-blackness and data colonialism. To center SSA voices in the global data ethics discourse, we present a framework for African data ethics informed by the thematic analysis of an interdisciplinary corpus of 47 documents. Our framework features six major principles: 1) Challenge Power Asymmetries, 2) Assert Data Self-Determination, 3) Invest in Local Data Institutions & Infrastructures, 4) Utilize Communalist Practices, 5) Center Communities on the Margins, and 6) Uphold Common Good. We compare our framework to seven particularist data ethics frameworks to find similar conceptual coverage but diverging interpretations of shared values. Finally, we discuss how African data ethics demonstrates the operational value of data ethics frameworks. Our framework highlights Sub-Saharan Africa as a pivotal site of responsible data science by promoting the practice of communalism, self-determination, and cultural preservation.

Beyond the Conversation: Designing the Architecture Around AI in Mental Health, Jina Suh ↗

Abstract: General-purpose conversational AI has become so embedded in everyday life that it works like a utility, and it now shapes how millions of people experience distress. Yet we still evaluate it like a product, one conversation at a time. This talk draws on surveys, interviews, and lived-experience accounts from people who turned to AI in distress. It looks at what happens around the conversation, following the whole journey: why people came to AI in the first place, what happened during the interaction, and what they did afterward. A safety redirect can score well on a benchmark and still cause harm, and people whom AI referred to a helpline rarely called it. The talk then looks at the people who carry the consequences of each safety feature: clinicians, crisis lines, families, communities, and the reviewers behind the scenes. It also covers the value tensions any design has to work through, such as intervention versus autonomy. I argue that the design problem is architectural. It is about who is accountable, what gets measured, and what paths a person has to safety. Standards should apply to that architecture: consent given in advance, multiple paths to safety, and measuring the whole journey. The responses themselves should stay tailored to the person.

Introduction and Overview, Abhishek Gupta

From Policy Models to World Action Models, Ge Yan ↗

Abstract: Robot learning has long relied on policy models that map observations directly to actions, while world action models (WAMs) that jointly predict future observations and actions are emerging as a powerful alternative. In this talk, I will trace this shift through my work, from ManiFlow, a flow-matching manipulation policy, and UniFlow, a unified 2D and 3D multi-modal policy, to Flex-π, a multi-stream world action model that jointly predicts future video, DINO semantics, 3D point maps, and actions. A single Flex-π model can run as a fast action-only policy, a full world action model, or anything in between at inference time, and it more than doubles the real-world success rate of the strongest baseline on bimanual, dexterous, and high-precision manipulation tasks.

Biography: Ge Yan is a Ph.D. student at the University of Washington, advised by Dieter Fox, and works with the Toyota Research Institute. His research focuses on robot learning, in particular multi-modal policy and world action models that combine video, semantic, and 3D representations for dexterous manipulation. His recent work includes ManiFlow, UniFlow, and Flex-π. 

Large-Scale Reinforcement Learning in Simulation for Dexterous Robotic Assembly, Patrick Yin ↗

Abstract: Robot learning needs large, diverse data to acquire dexterous, general-purpose skills. Physics simulation can generate that data at scale, but two central challenges remain. The first is generating meaningful behaviors in simulation without extensive human effort. The second is deciding what to train on as task complexity and the number of tasks grow. This talk presents three works that build on each other to address these challenges. OmniReset makes long-horizon reinforcement learning tractable by training from a broad distribution of reset states with a single reward function. It removes the need for demonstrations, curricula, and reward tuning, producing skills such as screwing, insertion, and assembly that transfer to real hardware from RGB cameras. SGS (Success-Guided Sampling) adaptively focuses training on the tasks and conditions where the policy is still learning. AlphaNIST combines these ideas to train a single policy on all 20 tasks of the NIST assembly board using only binary task rewards. It reaches 98.2% overall success and outperforms specialists trained on each task separately. Together, these results suggest that binary task rewards and multi-task training, both historically harder, become advantages at scale.

Biography: Patrick Yin is a Ph.D. student at the University of Washington’s Paul G. Allen School of Computer Science & Engineering, advised by Abhishek Gupta, and an NSF Graduate Research Fellow. He is also a research intern at NVIDIA. His research focuses on sim-to-real reinforcement learning for contact-rich robotic manipulation. Previously, he interned at Microsoft Research and completed his undergraduate studies at UC Berkeley, where he worked with Sergey Levine.

Cyber-Mechanical Intelligence: Creating Near-Gram Robots That Can Move Autonomously, Kyle Johnson

Abstract: Near-gram robots capable of autonomous movement could transform distributed sensing, environmental monitoring, and inspection in environments inaccessible to larger platforms, yet achieving true autonomy at sub-gram scales demands a fundamentally different design philosophy. This talk introduces Cyber-Mechanical Intelligence (CMI), a co-design paradigm that treats the location of intelligence within a robot as a design variable, distributing autonomous function across physical, electronic, and computational domains to meet task requirements while minimizing system-level resource cost. Three robotic systems are presented as demonstrations of CMI in practice. The Origami Microflier (414 mg) encodes aerodynamic decision-making into a bistable origami structure, enabling solar-triggered mid-air morphological reconfiguration and programmable dispersal of up to 98 m without continuous power. The MilliMobile achieves battery-free terrestrial locomotion from ambient solar and RF energy as low as 50 µW, and extends to onboard visual navigation through a lightweight embedded vision pipeline. The Coin-Copter offloads roll and pitch stabilization to passive flybar-rotor dynamics, reducing the sensing and control burden of a 1 g aerial vehicle while demonstrating closed-loop yaw control and a lift-to-weight ratio of 4.2. Together, these systems argue that intelligent behavior in resource-constrained robots is best understood as a property of the complete cyber-mechanical system, not of computation alone.

Introduction and Overview, Linda Shapiro

Online Neural Space Time Memory for Dynamic Novel View Synthesis, Baback Elmieh ↗

Abstract: Online novel view synthesis from multi-view streaming videos faces a fundamental trade-off: maintaining a persistent, long-horizon memory to reconstruct temporarily occluded regions while operating under strict real-time constraints. While Test-Time Training (TTT) offers a powerful memory mechanism, standard models mandate gradient-based memory updates at every frame to adapt to the changing motion in dynamic scenes. The computational cost of heavy memory updates precludes real-time application and can lead to instability over long contexts. Given that memory updates are more demanding than memory application and video content is largely redundant, we propose to decouple the frequencies of these two processes. Our approach performs periodic memory updates while applying the memory on a per-frame basis, using cross-view attention to manage deformations between the prior memory state and the current frame. To lock in the historical context, we introduce two critical mechanisms: an auxiliary Memory Loss that forces persistent internalization of the scene, and a Memory Caching strategy that regularizes active weights against catastrophic drift. Our method demonstrates state-of-the-art minute-scale memory persistence in online dynamic human scenes at amortized real-time speed. Visit webpage ↗

Biography: Baback Elmieh is a Computer Science PhD student at the University of Washington, where his research focuses on generative video models, real-time interactive systems, and persistent visual memory for dynamic 3D scenes. His recent work on Neural Space-Time Memory addresses a key challenge in immersive 3D telepresence: synthesizing consistent views of moving people by remembering visual details that are no longer visible to the cameras. Before returning to academia, he spent over two decades developing video games, real-time 3D rendering technologies, and AR/VR systems at Google, Meta, Qualcomm, and several startups. His work on Google’s Spotlight Stories earned an Academy Award nomination and an Emmy for innovation in interactive storytelling.

Toward Interactive Multimodal Agents, Zixian Ma ↗

Abstract: Multimodal agents have the potential to become a new interface through which people interact with digital and physical environments. While recent advances have enabled agents to perceive complex visual scenes and perform increasingly sophisticated tasks, building agents that can interact effectively with people and dynamic environments remains an open challenge. In this talk, I will present my research toward interactive multimodal agents, progressing from perception to action to interaction. I will first introduce Molmo2, a multimodal foundation model for understanding and grounding visual information. I will then discuss how multimodal capabilities extend to tool use and autonomous action through LATTE and MolmoWeb, which enable agents to interact with tools and navigate websites visually. Finally, I will introduce our ongoing work on Interactive Visual Gym, an environment for training and evaluating agents that interact with humans in dynamic tasks such as cooking and navigation. Together, these works explore how to move beyond multimodal understanding and autonomous task execution toward agents that can continuously perceive, act, and collaborate with people.

Biography: Hi! I am a 4th-year Ph.D. student in Computer Science and Engineering at the University of Washington, advised by Prof. Ranjay Krishna ↗ and Prof. Dan Weld ↗, and supported by the Qualcomm Innovation Fellowship ↗ and Two Sigma PhD Fellowship ↗. I build open multimodal models (Molmo2 ↗, MolmoPoint ↗), agents that act on the web and with tools (MolmoWeb ↗, LATTE ↗), and benchmarks of what these systems can and can’t do (Task Me Anything ↗, CREPE ↗). Currently, I’m especially interested in real-time, interactive multimodal models and agents.

Do Medical Language Models Preserve Correct Judgement Under Misleading Context?, Simon Zou ↗

Abstract: Clinical and consumer-facing deployments of large language models (LLMs) can now achieve expert-level scores on medical question-answering benchmarks based on licensing exams. However, real-world questions from users rarely resemble the clean, single-turn exam questions in these benchmarks. In reality, interactions are often layered with patient-reported history, retrieved documents, and false claims. This thesis asks whether an LLM that answers a medical question correctly in a clean environment continues to be correct once such misleading context is added. In short, it often does not. To study this question at scale, this thesis presents MedMisBench, a benchmark of 10,932 medical question items paired with 48,889 misleading context option sentences spanning medical-reasoning, agentic-capability, and patient-journey question formats. Each injected sentence is constructed along two independent axes that separately capture the type of medical false claim and its apparent source. This way, failures can be attributed to what is false and who asserts it. Evaluated across 11 configurations of commercial, open-source, and medical-domain LLMs, mean answer accuracy falls from 71.1% on clean questions to 38.0% once an injection sentence is prefixed to one answer option, with a mean attack success rate of 51.5%. Fabricated treatment exceptions and claims attributed to clinical authorities are consistently the most effective types in lowering the accuracy of LLMs, with attack success rates as high as 64.1% and 69.5% respectively. A panel of 14 clinical reviewers spanning 7 countries judged that 38.2% of these injection-induced errors carried serious potential harm to a patient. These findings indicate that strong performance on existing medical benchmarks does not mean that a model will preserve sound clinical judgment once its context is contaminated by misleading information. By releasing MedMisBench as a static, reusable benchmark, this work provides a foundation for evaluating and eventually mitigating this gap between benchmark performance and real-world robustness.

Biography: I’m a first-year Ph.D. student in Electrical and Computer Engineering here at UW, advised by Professor Linda Shapiro. My research focuses primarily on medical applications of computer vision and AI, from evaluating whether medical language models hold on to correct answers under misleading context to building egocentric vision systems that automate welfare checks for laboratory animals. I’m co-first author of MedMisBench, the NeurIPS 2026 benchmark paper I’m presenting today. I hold an M.S. from the University of Washington and a B.A.Sc. in Systems Design Engineering from the University of Waterloo.

Measurement-Conditioned Virtual Try-On, Yuanhao Wang ↗

Abstract: Real-world clothing try-on serves two purposes: visualizing how a garment looks on the body, and seeing how a specific size actually fits—whether it stretches tightly or drapes loosely. Current Virtual Try-On (VTO) models address appearance but ignore fit; conditioned only on 2D images, they default to synthesizing deceptively well-fitted results regardless of garment dimensions. In this work, we present RealFIT, the first scalable framework for real-world, measurement-conditioned virtual try-on across complex apparel categories. To overcome the scarcity of metric-labeled data and the sim-to-real gap of 3D physics engines, we introduce a multi-tier data hierarchy combining web retail metadata, VLM-assisted annotations, and a newly curated real-world dataset with tape-measured ground truth. We further propose a flow-matching VTO architecture featuring a dedicated measurement encoder and a Localized Metric Contrastive (LMC) loss that contrasts velocity fields across sizing perturbations to sharpen fit sensitivity. Finally, we establish a systematic evaluation protocol and demonstrate that RealFIT achieves state-of-the-art photorealism, fit accuracy, and size controllability.

Biography: Yuanhao is a second-year PhD student in the Graphics and Imaging Lab at UW, co-advised by Prof. Ira Kemelmacher-Shlizerman and Prof. Steve Seitz. He is also a student researcher at Google. Before UW, he was a master’s student at Carnegie Mellon University advised by Prof. Fernando De la Torre, and an undergraduate student at Brown University advised by Prof. James Tompkin. His research focuses on computer vision and generative models.

Introduction and Overview, Franzi Roesner

View the Grand Challenge that relates to this topic

Online Safety in the Age of AI, Grace Brigham ↗

Abstract: As AI technologies proliferate online, they are reshaping the online safety landscape by creating new ecosystems and disrupting existing ones. Understanding these impacts and developing effective responses is a pressing priority for security research. In our work, we employ human-centered methods to address this challenge, with a specific focus on at-risk users.

Reverse Engineering the Data Mixture, Sewoong Oh

Abstract: We are living in a data-centric AI era where data is the key in making frontier models. One secret that frontier labs guard fiercely is data mixture: how much data from each source should we mix in to create the pile of data for training the model? We reverse engineer this data mixture for proprietary models, using information from their tokenizers.

Benchmarking and Defending Against AI-Assisted Bioweapon Development, Henry Wong ↗

Abstract: AI-assisted bioweapons development has become a recent concern due to the rise of biological design tools such as ProteinMPNN and Evo2, along with advancements in automated laboratory instrumentation such as benchtop DNA synthesizers and cloud lab innovations. This talk will lay out specific cyberbiosecurity threat models enabled by AI and automation, and present our recent work in benchmarking and mitigation.

Privacy-Preserving Authentication of Digital Identity, Champ Chairattana-apirom ↗

Abstract: Digital identification and authentication are becoming increasingly important in our everyday lives. From new age-verification legislation, the European Union’s push for digital identity wallets, to the growing use of AI agents to interact with online services, these developments raise important questions and concerns about user privacy. In this talk, we discuss how cryptographic tools can help address these privacy concerns, enabling users to authenticate themselves without revealing unnecessary personal information.

Introduction and Overview, Vikram Iyer

View the Grand Challenge that relates to this topic

Accelerating Sustainable Computing Design with AI, Zhihan Zhang ↗

AI is becoming ubiquitous, embedded in the everyday devices and tools we use to understand and build the world, but its environmental footprint is surging just as fast. Achieving meaningful reductions in climate impact requires reimagining how we design and assess the carbon footprint of computing systems. In this talk, I will first highlight key challenges in incorporating sustainability considerations like carbon emissions when designing these technologies. I will then present a multimodal AI system that collects and reasons about publicly available data to reduce cradle-to-gate life-cycle assessments from weeks to seconds. Finally, I will show how AI can go beyond assessment to engineer better solutions: using feedback from real hardware, AI iteratively optimizes physical AI systems for metrics such as power, outperforming human experts and compressing models by orders of magnitude.

Vision foundation models for species-agnostic behavior analysis, Moses Lurbur ↗

Abstract: Generalizable, accessible tools for quantifying and classifying animal behavior in videos are crucial for conservation, biology, and medicine. Across the rich suite of tools for analyzing animal behavior, most rely on keypoint detection models trained on species-specific, experiment-specific, manually annotated data. To address this limitation, we have developed a species-agnostic video processing methodology that exploits the emergent properties of object segmentation and image-embedding foundation models. To make our methods accessible to scientists, we built DETAIL. With DETAIL, users can discover, label, and analyze detailed behaviors on anything from mice to snow crabs. All with zero manual annotation.

Power-Adaptive Datacenter Design, Jonggyu Park ↗

Abstract: The rapid growth of AI and cloud computing is driving unprecedented demand for datacenter power. At the same time, fluctuating workloads and increasingly dynamic power availability pose new challenges for efficient datacenter operation. In this talk, I will present my research on power-adaptive datacenter design, focusing on predictive power management for general-purpose compute servers. I will introduce PIP (Prediction-Informed Power Management), which predicts the power impact of resource configurations before applying them, enabling rapid adaptation to changing power budgets while maximizing power utilization and application performance. I will also briefly discuss my work on power-adaptive storage and elastic memory systems, and conclude with my vision for full-stack power-adaptive datacenters that coordinate compute, storage, and memory resources under dynamic power constraints.

Towards Efficient and Context Aware Biodiversity Vision Systems for Edge Deployment, Emmanuel Azuh Mensah ↗

Abstract: Deploying vision models on wildlife camera traps exposes three compounding failure modes. First, efficient mobile architectures are static: their capacity cannot adapt to the high input variability of camera trap imagery. Second, compact models lack the taxonomic knowledge encoded in large vision-language foundation models such as BioCLIP2, yet those models are too large for on-device inference. Third, edge-deployed vision systems typically discard geographic location, a strong prior over species identity available in most camera trap deployment. In this talk, I’ll discuss what makes distilling foundation models into compact conditional-adaptive vision models challenging and where external context and knowledge of the teacher models’ encoded geometry show promise towards more accurate, efficient deployable systems. Together, this work builds toward an adaptive, knowledgeable, and context-aware vision system deployable in the wild without connectivity with models easily accessible on huggingface and deployed using popular edge devices such as Google Edge TPU and Apple Neural Engine.

Introduction and Overview, Magda Balazinska

R3Con: Context Representation for Complex Question Answering over Large Corpora, Michael Theologitis ↗

Abstract: Coming soon

KathDB-FAO: Efficient Question Answering over Multimodal Data via Code Generation and Optimization, Guorui Xiao ↗

Abstract: Coming soon

How People Debug Complex SQL with LLMs: A Qualitative Study, Nicole Sullivan ↗

Abstract: Coming soon

Introduction and Overview, Jason Hoffman

3:55-4:15: EvoClinician: Self-Evolving System for White-Box Multimodal Clinical Reasoning, Yongkang Li

Abstract: Coming soon

Dissecting and Directing Pathology Foundation Models, Chanwoo Kim

Abstract: Foundation models (FMs) are central to recent advances in digital pathology, encoding histology images into dense embeddings for diagnostic classification, molecular alteration prediction, and clinical outcome modeling. However, the opacity of these embeddings renders such systems “black boxes,” limiting their clinical translation and scientific utility. Here, we introduce PICASSO (Pathology Image Concept Atlas built via SparSe dictiOnary learning), a framework that makes pathology FMs interpretable and controllable. It decomposes FM embeddings into human-interpretable visual concepts using a sparse autoencoder, a form of dictionary learning. Trained on more than 120 million tissue patches across 32 cancer types, PICASSO produces the first pan-cancer atlas of histomorphological concepts. Beyond interpretation, PICASSO enables concept editing, the targeted manipulation of concepts to interrogate and control model behavior. These capabilities enable diverse applications, including model auditing, biological discovery, and in silico interventions. PICASSO transforms FMs from black-box predictors into interpretable and controllable models that support biological insight, hypothesis generation, and virtual experimentation.

PeriPulse: Open-Ear Headphones for Robust Heart Rate Sensing During High-Intensity Exercise, Cynthia Dong

Abstract: Live cardiovascular monitoring has potential to unlock new interactive health experiences, but motion artifact noise and limited form factor options prevent users from accessing high-accuracy sensing. Leveraging the popularity of open-ear headphones and their unique contact pressure, advantageous physiological positioning, and potential for audio interaction, we present \device, an accurate, motion-robust headset which integrates neatly into users’ existing habits. We iteratively designed PeriPulse based on in-lab studies, then collected and trained models on a high-intensity exercise dataset of 29 participants, extracting heart rate with an error of 8.11 +- 8.30 BPM. We then deploy models for real-time heart rate estimation and generate audio-based heart zone alerts with 9 participants. Users found the system accurate (error: 5.03 +- 2.16 BPM), appreciated verbal notifications over visual interfaces for zone training, and saw potential to consolidate other trackers into their headphones. Our results illustrate \device’s capacity for motion-robust sensing, and suggest opportunities for innovative end-user applications.

Flex-π: A Multi-Stream World-Action Model with Compute Flexibility
Authors: Ge Yan, Jinghao Liu, Yuzhi Fan, Lei Cai, Minwen Liao, Jesse Zhang, Dieter Fox
Presenter: Ge Yan ↗
Lab/Research Area: RSE Lab, Robotics

Abstract: Developed Flex-π, a world-action model that jointly predicts future appearance, 3D geometry, and object semantics in one shared latent space, addressing a key limitation of existing world models that rely on RGB video alone. A single checkpoint runs either as a fast action-only policy or as a full world model, letting users trade latency for accuracy at deployment without retraining. On a real bimanual robot, it outperforms the strongest baselines by 2–6× on precise, dexterous tasks, including a robot repairing its own gripper with tight-margin insertions, and the full code, data, and models are open-sourced.

When Agents Slow Down: Understanding LLM Agents’ Test-Time Strategies via Elo-per-token Analysis
Authors: Kaiyuan Liu, Qiuyang Mang, Bo Peng, Wenhao Chai, Hanchen Li, Shreyas Pimpalgaonkar, Luke Zettlemoyer, Alex Dimakis, Alvin Cheung
Presenter: Kaiyuan Liu ↗
Lab/Research Area: ZLab

Abstract: Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for intermediate submissions, making progress observable throughout long trajectories. We propose Elo-per-token analysis, which tracks the best solution found at each token budget and uses a Bradley-Terry model to aggregate within-task orderings into Elo ratings across tasks with different score scales. We apply it to four general-purpose agents on four open-ended benchmarks, with sessions of up to 100M tokens, and to three feedback-driven LLM optimization harnesses in controlled single-task interventions. Independent sampling provides a theoretically characterized reference, for which Elo grows linearly with log compute. Against this reference, agents can initially convert tokens into Elo faster than independent sampling, but their marginal gains diminish and eventually fall below the reference. In contrast, the strongest historical human contestants improve superlinearly over contest time on shared AtCoder Heuristic Contest tasks, providing evidence of continual learning and substantial headroom after agents slow down. We define the scaling inflection point as the per-session budget where marginal Elo gains match the independent-sampling reference. Using this point as the per-session budget, we split 100M tokens across parallel sessions on FrontierCS Polyomino Packing, gaining +264 Elo over one long session and +355 over ten short sessions.

Forty Shades of Blue: Quality-Diversity Alignment via Mode-Conditioned Reinforcement Learning
Authors: Jiayi Yuan, Hangoo Kang, James Jihao Liu, Yejin Choi, Vikram Iyer, Liwei Jiang, Natasha Jaques
Presenter: Jiayi Yuan ↗
Lab/Research Area: SocialRL Lab

Abstract: We present MoDA (Mode-conditioned Diversity Alignment), an online post-training RL algorithm that jointly optimizes generation quality and diversity, inspired by the coordination perspective in multi-agent reinforcement learning (MARL). MoDA trains a single shared LLM policy conditioned on abstract numbered roles, where each role acts as an agent competing to produce outputs distinct from the others. This formulation encourages mode-conditioned agents to explore complementary regions of the high-quality output space without requiring hand-crafted personas or architectural modifications.

AutoSciency: Mapping Knowledge Gaps in Embedding Space
Authors: Jiayi Yuan, Aarav Talati, Emma Huang, Sota Kobata
Presenter: Jiayi Yuan ↗
Lab/Research Area: SocialRL

Abstract: Prior work shows that unexpected combinations of content and context can predict impact, while some embedding methods provide scientific document embeddings for downstream tasks [4]. Our project builds on these ideas by asking whether gaps in embedding space can be translated into human-readable research hypotheses.

Narrative in Agentic Coding
Authors: Steve Tanimoto
Presenter: Steve Tanimoto ↗
Lab/Research Area: HCI, AI

Abstract: As gen-AI agents become more powerful, the gap between human understanding and what agents are doing is widening. This research explores one means to bridge that gap.

A Few Words Go a Long Way: Language Guided Robot Policy Synthesis
Authors: Daphne Chen, Archit Ritesh Jain, Eric Goossen, Emma Romig, Michael Murray, Nick Walker, Maya Cakmak
Presenter: Daphne Chen ↗
Lab/Research Area: Robotics

Abstract: While vision-language-action models have demonstrated impressive zero-shot manipulation capabilities, they remain fundamentally black box policies that are difficult to interpret, adapt, or correct when they inevitably fail. In this work, we propose ARCHITECT, a framework that treats robot policy acquisition as an interactive program synthesis task. ARCHITECT leverages the reasoning capabilities of LLM coding agents to synthesize modular robot programs that utilize a suite of perception and control tools. Unlike end-to-end models where distribution shift leads to unpredictable, cascading failures, our modular architecture allows users to isolate failures and localize feedback at the level of abstraction required. We introduce an iterative process where a human supervisor provides natural language corrections to steer the policy. These corrections are grounded in the policy code by program execution traces and distilled into a persistent skill library, enabling long-term in-context learning which allows the agent to accumulate a repertoire of reusable, interpretable behaviors. In a benchmark evaluation on a Franka Panda robot, ARCHITECT outperforms state-of-the-art VLA models and program synthesis baselines on complex, long-horizon tasks, including articulated object manipulation and cloth folding. Our results demonstrate that the synthesized skill library enables the system to transfer to novel tasks with decreasing human intervention, providing a steerable and data-efficient alternative to
black-box robot learning.

OpenDubs: A low cost mobile manipulation robotics research platform
Authors: Emma Romig, Daphne Chen, Sidharth Talia, Personal Robotics Lab
Presenter: Sidharth Talia ↗, Daphne Chen ↗, Emma Romig ↗
Lab/Research Area: Robotics

Abstract: OpenDubs is an open-source mobile manipulator built at the University of Washington to give students a research platform with much of the functional envelope of a commercial humanoid: holonomic mobility, two 7-DOF arms, a vertical lift, an expressive head, and GPU autonomy compute. It costs a fraction as much, and is built mostly from off-the-shelf parts. The robot is built around an NVIDIA Jetson AGX Thor.

Synth∀ Instruction-Centric Microarchitectural Side Channel Mitigations
Authors: Alexandra E. Michael, Sushant Dinesh, Antonio Ballesteros, Dan Grossman, Christopher W. Fletcher, David Kohlbrenner
Presenter: Alexandra E. Michael ↗
Lab/Research Area: Security & Privacy / PLSE

Abstract: As Moore’s Law slows down, hardware manufacturers continue building novel microarchitectural optimizations into their chips. These optimizations increasingly create exploitable side channel vulnerabilities, yet the availability of defenses has not kept pace. Hardware- and software-level mitigations alike require time and expertise to develop—time in which a vulnerability remains open, and expertise that is often in short supply. We address these challenges by introducing (1) leakage equivalence classes, a lightweight and generalizable approach to formally specifying instruction-centric leakage; and (2) Synth∀, a novel program synthesis framework that uses such specifications to automatically generate safe instruction transformations for x86_64

Synth∀ combines syntax-guided synthesis, used to generate safe dataflow sketches, with superoptimization, which automatically refines sketches into verified, targeted transforms, all with minimal expert input. We apply Synth∀ to two broad classes of microarchitectural optimizations, computation simplification and significance compression, that impact traditionally constant-time instructions. Automated synthesis for the most complex instructions remains a challenge; yet Synth∀ provides transforms for a broad set of commonly used instructions across both case studies, while requiring little or no expert guidance beyond the initial leakage specification. Furthermore, the resulting mitigations are often shorter and more novel than the previous state of the art.

Can Vision-Language Models Analyze Human-Centered Videos? Mapping Model Capabilities and Human-AI Collaborative Workflows
Authors: Xiyuan Shen, Jiuyang Lyu, Seokhyun Hwang, Huanfen Yao, Shwetak Patel, Zhihan Zhang, Jacob O. Wobbrock
Presenter: Xiyuan Shen ↗
Lab/Research Area: Ubicomp Lab; HCI

Abstract: Video captures human behavior and interaction, but annotation remains labor-intensive and requires project-specific judgments. As Vision-Language Models (VLMs) become more capable, when can they annotate independently, and when does reliable analysis still require human involvement? We compared annotation workflows across 15 representative tasks. VLM-alone annotation approached human accuracy on average (HNS = 97.0). Human verification of VLM outputs achieved the highest mean accuracy (HNS = 121.5), while reducing 48.9% human annotation time and 31.3%–44.5% estimated cost than human-alone annotation.

AI-generated characterization of landscape risk for disease emergence Washington
Authors: Poojita Garg, Jared Palazza, Girish Narayanswamy, Peter Rabinowitz, John Choe, Shwetak Patel, Julianne Meisner
Presenter: Poojita Garg ↗
Lab/Research Area: Ubicomp Lab, Ubiquitous Computing

Abstract: Pathogen emergence accelerates at human–wildlife interfaces, where ecological change, host communities, and human activity converge. However, emergence events are rare or rarely detected limiting high-resolution prediction models trained directly on event data.
We use epidemiologic and ecological theory to model spillover risk factors rather than rare emergence events directly. We develop spatially explicit, high-resolution proxies for:
– Land-use structure
– Host-community composition
– Human–wildlife contact intensity

KathDB-FAO: Synthesized Query Plans in a Multimodal DBMS
Authors: Guorui Xiao, Douglas Brown, Artur Borycki, Magdalena Balazinska
Presenter: Guorui Xiao ↗
Lab/Research Area: UWDB, Data Management & AI

Abstract: Recent advances in large language models (LLMs) and vision-language models (VLMs) allow data systems to answer analytical queries over text and images alongside relational tables, but at a significant cost for large-scale data. Existing systems reduce this cost through rewrite rules written in advance, which cannot use the logic of an individual query or the semantics of the data it processes. To address this challenge, we design, implement, and evaluate KATHDB-FAO, an efficient query evaluation system over multimodal data. KATHDB-FAO first decomposes a complex query into an execution plan of fine-grained steps, which we call Function-as-Operators (FAOs). To reduce execution cost, it selects subsets of these FAOs to implement together as one function, enabling query-specific optimizations that make fewer model calls. Synthesizing these functions with a frontier model has its own cost, yet we observe that many synthesized FAOs recur across queries. Thus, to further reduce plan generation cost, KATHDB-FAO reuses the FAOs synthesized for earlier queries in a workload and adapts them to each new query. On SemBench and an additional workload, KATHDB-FAO achieves quality on par with the strongest baseline at an end-to-end cost per query that is 52% lower on average and up to 73% lower.

PartEinsum
Authors: Jiacheng Wu, Kyle Deeds, Dan Suciu
Presenter: Jiacheng Wu ↗
Lab/Research Area: UWDB

Abstract: To make einsum computations feasible under a given memory budget, we partition both the input tensors and the computation steps. We also introduce a cost model that estimates computational cost and propose algorithms that use it to find better execution plans.

How People Debug Complex SQL With LLMs: A Qualitative Study
Authors: Nicole Sullivan, Ella Cao, Leilani Battle, Magda Balazinska
Presenter: Nicole Sullivan ↗
Lab/Research Area: Database Group

Abstract: We lack a fundamental understanding of how LLMs impact the SQL query debugging workflow. In response, we conducted a user study with 16 participants of varying SQL expertise, examining how they debug complex SQL queries with and without LLM support. Qualitative analyses of our study data revealed that while participants trusted the LLM, they varied widely in how and whether they delegated debugging tasks to the LLMs. Additionally, participants enforced specific collaborative roles for the LLM such as TA or sounding board, which participants controlled by withholding context from the LLM. These findings suggest that LLM-based tools could enable preset role selection, broader context management, and flexible automation features, including low automation options, to support SQL debugging moving forward.

FlexiCodes: Embedded De Bruijn Patterns for Deformation Tracking on Elastic 3D-Printed Objects with a Monocular Camera
Authors: Daniel Campos Zamora, Raiden Santos*, Ria Agarwal*, Jerry Cao, Jon Froehlich, Chang Xiao
Presenter: Daniel Campos Zamora ↗
Lab/Research Area: Makeability Lab, HCI

Abstract: Non-rigid interfaces allow for soft, deformable inputs using malleable materials, which enable interactions that would be infeasible with traditional rigid controllers. However, tracking and decoding the deformations typically requires embedded sensing hardware or complex external setups (e.g., stereo cameras, calibrated projector setups). We present FlexiCodes a novel approach to deformable interaction that leverages 3D printing to embed De Bruijn patterns onto elastic objects and computer vision to track surface deformation. Our three-part system consists of design, fabrication, and computer vision components that allow for vast design possibilities for flat and volumetric forms, preserve the ability to tune mechanical properties, and enables live interactions, respectively. We perform a technical evaluation and validate our approach through a series of proof-by-demonstration scenarios.

Thinking while Speaking: Inference-Time Knowledge Transfer for Responsive and Intelligent Conversational Voice Agents
Authors: Zachary Englhardt ↗
Presenter: Zachary Englhardt, Justin Chae, Bilal Bhurgri
Lab/Research Area: Ubicomp Lab

Abstract: Voice agents face a fundamental tension. The reasoning, retrieval, and tool use that make foundation models capable are iterative and slow, while conversation demands responses on a millisecond timescale. Small real-time models meet the latency bar but cannot match foundation models on complex tasks, leaving current voice agents to trade away either responsiveness or capability. We introduce conversational infill, in which a small Talker model both immediately generates contextually grounded responses to hide the latency of an external Reasoner model and fluently integrates streamed Reasoner knowledge into its responses during inference. We curate a 290,571-example synthetic dataset spanning six domains and show that the task is learnable across seven widely used small language models ranging from 135M to 1.7B parameters. Our system, ConvFill, sustains millisecond-level time-to-first-response while closing the accuracy gap to within 6.3% of the corresponding frontier Reasoner. In a live user study on Apple M2 hardware (n=18), participants rated ConvFill on par with frontier models overall, preferred it for retrieval tasks, and ranked it significantly more responsive.

GazeRL: Cognitive Reward Shaping for Scanpath Prediction
Authors: Jiexin Ding, Xinyun Liu, Ruotong Yu, Zhihan Zhang, Shwetak Patel, Yue Jiang, Akshay Gadre
Presenter: Jiexin Ding ↗
Lab/Research Area: Ubicomp Lab

Abstract: GazeRL fine-tunes a vision-language model (Qwen2.5-VL-3B) to predict human visual-search scanpaths using reinforcement learning with cognitively grounded rewards: decaying bottom-up saliency, ramping top-down semantics, and object-level inhibition of return. On COCO-Search18, it produces scanpaths that are both accurate and human-like diverse, ranking best or second-best on every metric that survives our audit.

Single-molecule recording of protein age using a delamination based molecular clock
Authors: Mattias Tolhurst, Daphne Kontogiorgos-Heintz, Safwan Diwan, Jeff Nivala
Presenter: Mattias Tolhurst ↗
Lab/Research Area: MISL – Synthetic Biology, Biotechnology, Biochemistry

Abstract: Proteins are the functional unit of biology, and despite often being treated as static once expressed, they are in fact regulated in complicated ways. For example, proteins acquire post-translational modifications (PTMs), interact within the cellular milieu, and acquire chemical modifications as they age. In addition, their expression, and therefore their interactions differ between tissues, cell types, and subcellular compartments. Despite this known complexity, the current tools that biologists use to measure protein age and turnover do not operate at the granularity required to capture this heterogeneity. In this work, we propose using a small, genetically encoded tag, with predefined deamidation regions to measure protein age. We demonstrate that not only does the deamidation correspond closely to protein age in vivo, but also that it may be possible to reconstruct protein age on a single molecule level, the first such technique capable of doing so.

Enabling Minimally Perturbative Nanopore Protein Sequencing through Modular Adapter Conjugation
Authors: Kingsley Wong, Mattias Tolhurst, Ron Shanderson, Daphne Kontogiorgos-Heintz, Lacramioarra Bintu, Jeff Nivala
Presenter: Guorui Xiao ↗
Lab/Research Area: Molecular Information Systems Laboratory (MISL)

Abstract: Nanopore-based protein sequencing offers a promising approach for single-molecule protein analysis, but its broader applicability may be limited by the need to fuse target proteins to bulky sequencing adapters that can interfere with native protein function. This project develops a modular adapter conjugation strategy to separate target protein expression from sequencing adapter incorporation, enabling adapters to be attached after protein production and prior to nanopore sequencing. By decoupling these steps, the approach aims to preserve target protein functionality while meeting the structural requirements of the ClpX-mediated nanopore sequencing platform. This strategy provides a foundation for more flexible library preparation and may facilitate the application of nanopore protein sequencing to a wider range of biologically relevant proteins.

CapSleeve: A Smart Rehabilitation Tool
Authors: Abhinav Komanduri, Richard Li, Alexander H. Nelson, Shwetak Patel
Presenter: Abhinav Komanduri ↗
Lab/Research Area: Ubicomp Lab

Abstract: Physical rehabilitation depends on frequent clinic visits and guided feedback that many patients can’t easily access. CapSleeve is a low-cost textile elbow sleeve that fuses capacitive sensing with IMUs to track arm motion and muscle activation, and our preliminary prototype shows distinct signals for hand and muscle activity, a first step toward at-home rehab with automated session summaries.

AI Iteratively Optimizing AI via Real Hardware Feedback
Authors: Zhihan Zhang, Alexander Metzger, Jiuyang Lyu, Chun-Cheng Chang, Jiayi Shao, Yujia Liu, Emmanuel Azuh Mensah, Edward Wang, Kurtis Heimerl, Gregory D Abowd, Shwetak Patel, Natasha Jaques, Vikram Iyer
Presenter: Zhihan Zhang ↗
Lab/Research Area: Ubicomp, LLM, Physical AI

Abstract: As AI moves into the physical world, from robots to sensors to wearables, models increasingly must run on the device itself. A wearable cannot always send private health data to the cloud, and a camera trap in the wilderness may have no connection at all. Fitting a model onto a tiny edge device means simultaneously meeting hard limits on memory, power, and temperature while preserving accuracy. This is performed manually by experts today, and each new device or model means starting over. AI agents now write software well because code can be run and checked instantly. Physical systems are different. Their behavior depends on the real hardware, and simulations often fail, a problem robotics calls the sim-to-real gap. We found the same gap here: frontier models fail entirely without hardware feedback (0% deployment success).
To close this gap, we built a hardware-in-the-loop environment in which the agent iteratively refines both model and device firmware by measuring on real hardware. With this feedback, the agent surpasses human expert results within seven iterations. We demonstrate practical impact in two real-world deployments, developed in direct collaboration with domain practitioners: an elk-detection camera trap, with 250× compression for the vision model and under 3.3% accuracy loss; and a phonetic-transcription wearable for child development research, with 400× compression for the audio model and under 6% feature error rate loss.

ArtScape: Detecting, Characterizing, and Interactively Exploring Urban Art at Scale using Vision Language Models
Authors: Lucy Jiang, Jared Hwang, Jon E. Froehlich
Presenter: Jared Hwang ↗
Lab/Research Area: Makeability Lab, HCI

Abstract: Urban art brings cultural expression, community identity, and aesthetic enrichment to cities. Currently, there are no systematic ways of exploring urban art at scale. We introduce ArtScape, a VLM-based interactive tool that lets users view, search, filter, and tour four categories of art (murals, mosaics, sculptures, graffiti). To comprehensively and automatically detect urban artwork throughout a region, we designed a custom five-stage pipeline that uses VLMs to detect and describe urban artwork within 360° panoramic Street View images, achieving a micro-F1 of 0.83 on an urban art dataset we labeled. ArtScape categorizes, describes, and scores the reachability and interestingness of each artwork while deduplicating based on proximity and similarity. We evaluated this tool with 12 participants spanning urban planners, artists, sighted users, and blind and low vision users. Our findings show that ArtScape’s at-scale AI artwork curation enables granular and aggregate urban art discovery, thereby helping people learn about their communities and city-wide trends, motivating in-person urban exploration, and expanding perceptions of what constitutes urban art.

Realize What Matters: Principled Context Representation for Large-Scale Reasoning
Authors: Michael Theologitis, Dean Light, Shuyue Stella Li, Benjamin Newman, Yulia Tsvetkov, Dan Suciu
Presenter: Michael Theologitis ↗
Lab/Research Area: NLP, Data Management

Abstract: Solving complex tasks in domains such as science, medicine, law, and finance often requires assembling interdependent information scattered across vast, heterogeneous sources far beyond model context limits. Existing approaches tackle this challenge by organizing information into more manageable representations over which models can reason, such as graphs, textual memories, and retrieval collections. These representations dictate what downstream reasoning is possible and, ultimately, whether it succeeds; yet their design and construction remain largely ad hoc. In this work, drawing on the cognitive theory of relevance realization, we propose concrete principles for designing AI systems that construct effective representations of very large contexts. We analyze existing approaches and show how their successes and failures can be explained by their alignment with these principles, and introduce R3Con, a harness that systematically operationalizes them. We evaluate R3Con against nine state-of-the-art baselines on two recent benchmarks of reasoning over large document corpora. On these benchmarks, R3Con substantially outperforms the strongest baseline, by 20 and 8.4 percentage points. It also enables smaller models to outperform much larger ones: R3Con with 4B and 9B models outperforms all evaluated 35B baselines, while R3Con with Qwen3.5-35B-A3B outperforms Claude Code with Claude-Sonnet-5 at 3.7× lower cost. Our results show that context representations following our principled approach can reduce reliance on model scale, pointing toward a future of AI systems with frontier-level performance powered by smaller models.

Bicameral: Memory-Efficient Transformer via Latent Tensor Rematerialization
Authors: Zhongyi Kurt Gu, Vikram Iyer, Mark Oskin
Presenter: Kurt Gu ↗
Lab/Research Area: Computer Systems, Machine Learning

Abstract: Memory capacity limits the size of language models that can be trained and served. We introduce Latent Tensor Rematerialization (LTR), which generates each weight matrix on demand from compressed latent vectors with small learned generators, and in doing so converts a standard matrix into a recomputable activation. This reduces memory usage by directly storing fewer parameters and optimizer states, as an alternative to lowering precision. We validate the effectiveness of LTR with Bicameral, a LLaMA-style transformer architecture whose weight matrices are generated on demand by small decoders. Trained on 2B effective parameters to 40B tokens, Bicameral matches the uncompressed baseline on perplexity and zero-shot benchmarks, while reducing stored parameters by 50% and training memory usage by up to 48%.

On-device AI
Authors: Stefany Cruz, Junxi Xia, Franklin Li, Rahaf Alharbi, Vikram Iyer, Stephen Xia
Presenter: Stefany Cruz ↗
Lab/Research Area: Iyer Lab/ Ubicomp

Abstract: Co-designed with urban communities across four cohorts seeking better situational awareness against incidents of crime, Patronus is a body-worn system that identifies and continuously localizes sounds like approaching footsteps.

Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment
Authors: Soham Gadgil, Chris Lin, Su-In Lee
Presenter: Soham Gadgil ↗
Lab/Research Area: AIMS Lab, LLM Alignment

Abstract: Steering vectors have emerged as a lightweight and effective approach for aligning large language models (LLMs) at inference time, enabling modulation of model behaviors by shifting intermediate representations toward a target behavior. Many existing steering methods apply interventions at globally fixed locations in an LLM, effectively treating the optimal intervention location as input-invariant. However, representations relevant to a target behavior can be encoded at different LLM layers depending on the input. Theoretically, we show that different inputs may require steering at different layers to achieve alignment with a given behavior. Empirically, we observe that the most effective layer for steering varies substantially between inputs. Motivated by these observations, we introduce Where to Steer (W2S), a general framework that adaptively selects intervention layers conditioned on the inputs by learning a mapping from input representations to steering layers. Across multiple LLMs and alignment behaviors, W2S consistently outperforms fixed-layer baselines in both in-distribution and out-of-distribution settings. Our findings suggest that input-dependent layer selection is a key design dimension complementary to methods for steering vector extraction.

PeriPulse: Open-Ear Headphones for Robust Heart Rate Sensing During High-Intensity Exercise
Authors: Cynthia Dong, Liam Bakar, Andy Ye, Noel Prince Muthuplakal, Zimo Chen, Kurtis Heimerl, Shwetak Patel, Richard Li
Presenter: Cynthia Dong ↗
Lab/Research Area: Ubicomp

Abstract: Live cardiovascular monitoring has potential to unlock new interactive health experiences, but motion artifact noise and limited form factor options prevent users from accessing high-accuracy sensing. Leveraging the popularity of open-ear headphones and their unique contact pressure, advantageous physiological positioning, and potential for audio interaction, we present PeriPulse, an accurate, motion-robust headset which integrates neatly into users’ existing habits. We iteratively designed PeriPulse based on in-lab studies,
then collected and trained models on a high-intensity exercise dataset of 29 participants, extracting heart rate with an error of 8.11±8.30 BPM. We then deploy models for real-time heart rate estimation and generate audio-based heart zone alerts with 9 participants.
Users found the system accurate (error: 5.03±2.16 BPM), appreciated verbal notifications over visual interfaces for zone training, and saw potential to consolidate other trackers into their headphones. Our results illustrate PeriPulse’s capacity for motion-robust sensing,
and suggest opportunities for innovative end-user applications.

Misalignments in Generative Protein Design: Thermodynamic Filtering Fails to Predict Specificity
Authors: Safwan Diwan, Mattias Tolhurst, Jeff Nivala
Presenter: Safwan Diwan ↗
Lab/Research Area: MISL

Abstract: Generative protein design models frequently produce high-confidence binders in silico, yet wet-lab evaluation demonstrates that these binders often lack the expected biochemical specificity. To probe the boundary conditions of current design and evaluation pipelines, we present a pilot failure-mode analysis by designing binders for asparagine deamidation. This single-residue modification consists of an electrostatic charge shift that can act as a rigorous test of highly specific binder design. When evaluating our designs against this target using a custom in vivo split-spGFP biosensor, we identified a confounding pipeline misalignment. While binding was achieved after target optimization, binding specificity was inverted compared to Rosetta’s thermodynamic predictions. Curiously, retrospective cross-architecture analysis, using multiple folding models, revealed AlphaFold 3 correctly predicted the in vivo specificity natively. However, this prediction would have been overridden by the commonly used DDG thermodynamic metric used for evaluation. Ultimately, this case study demonstrates that appending standard thermodynamic filters to generative outputs can potentially degrade pipeline reliability. This exposes flaws both in generative models and in how post-hoc scoring weighs theoretical electrostatic rewards against energetic binding penalties.

Smart Suction Cups with Capacitive Tactile Sensing
Authors: Alex Choi, Le Minh Anh Nguyen, Paolo Torrado, Tim Amish, Sara Mouradian, Joshua Smith
Presenter: Alex Choi ↗
Lab/Research Area: Sensor Systems Lab, Robotic Manipulation

Abstract: Grasping in industrial warehouse environments can often lead to the occlusion of vision systems. To address this, we developed an add-on sensor system for commercially available suction cups which enables tactile sensing. We use this new sensing modality to improve the pick rate of a robot arm compared to a vision-only baseline.

SpatialSelect: Text-selective binaural audio generation from video, with explicit azimuth control
Authors: Guilin Hu, Shwetak Patel
Presenter: Guilin Hu ↗
Lab/Research Area: Ubicomp

Abstract: Video-to-audio models can now generate sound for a video, but they either render every source in a mono mix or produce binaural audio without letting the user pick which sound to generate. SpatialSelect does both. Given a video, a text prompt naming the target sound, and a requested direction, it generates binaural audio of only that sound, placed where the user wants it. SpatialSelect is a single flow-matching model with shared and per-ear transformer blocks.

Safe, Age-Appropriate GenAI for Children: Perspectives of Platforms, Parents & Children
Authors: Abhinaya S B, Meghna Gupta, Anupam Das, Jason Yip, Franziska Roesner
Presenter: Abhinaya S B ↗
Lab/Research Area: Security and Privacy Research Lab; Online Safety

Abstract: Children’s use of generative AI is rapidly rising, driven by easy access to general-purpose chatbots, a growing ecosystem of child-specific GenAI apps, and increasing adoption of GenAI in educational contexts. However, parents have limited control over their children’s consumption of AI content, with non-standardized parental controls across tools and vague content filters. We systematically study the GenAI parental control ecosystem, identify what aspects of children’s GenAI interactions parents want to control, and how children’s desires align or contrast with those of parents. We extend this work to evaluate how guardrail models can be used to enforce parents’ topic-specific restrictions for their children.

Low-cost, Reliable, Self-Evolving Data Agent for Data Ingestion Tasks
Authors: David Alexander, Johannes Gehrke, Hossein Ahmadi, Magdalena Balazinska
Presenter: David Alexander ↗
Lab/Research Area: Database

Abstract: Much of an organization’s data sits outside any database, in files of many formats. A coding agent can extract the tables from any such file, but it is expensive. A data engineer faced with the same problem might learn from extractions of previous files and be more efficient with future files. We ask how much freedom a data agent should get in learning and evolving itself to be more efficient in this data extraction task. We measure approaches on what it keeps as memory, how it organizes that memory, and when it updates it. On eight workloads with Claude Sonnet 5.5 and Opus 5.5, neither text nor code memory lowers cost much, because the agent still reads every file. An agent left to design its own harness lands anywhere from $5 to $210 on the same 200 files, with 39% to 65% of tables correct, which shows considerable variance. Proteus is a harness with a structured approach in evolution. It evolves learned extraction classes (a rule and a recipe each), a verifier that checks every extracted table, and a coder called only when no class matches or a check fails. Proteus matches the correctness of the no memory agent on every workload and is up to 6 times cheaper where many files share a class. Where nearly every file is its own class, writing reusable code (buying) every time costs more than writing one-off extraction (renting). We also investigate this buy vs. rent problem and create an approach which costs comparable to a rent-only agent on workload with few recurrences.

Symbolic Relations for Network Change Verification
Authors: Brody Franks (UW), Nicolas Koh (Princeton), Dan Wang (Xi’an Jiaotong University), Daniel Halperin (AWS), Zachary Kincaid (Princeton), Ratul Mahajan (UW), David Walker (Princeton)
Presenter: Brody Franks ↗
Lab/Research Area: Systems, PLSE, DB

Abstract: We introduce a new language and engine for network change verification. Our tool: (1) verifies both the control plane and the data plane, in one general formalism, (2) enables customization of the data representation at a low level of granularity, while retaining the same frontend specification, (3) supports specialized data structures that hybridize concrete and symbolic verification and are efficient on real-world workloads. We apply our system to representative network-change scenarios, examining both its expressiveness and performance, showing that it checks symbolic relational specifications 25-541× faster than a general-purpose solver.

Secure Screening for DNA Hardware
Authors: Henry Wong, Tadayoshi Kohno, Jeff Nivala
Presenter: Henry Wong ↗
Lab/Research Area: Security & Privacy, MISL

Abstract: DNA Benchtop Synthesizers allow consumer to print synthetic DNA. But what if a malicious actors tries to print a pathogenic sequence? Online screening on benchtops have been explored before by SecureDNA, but whether the screener is trustworthy and implementing a fail-off mechanism has yet to be explored. In my work I create a framework for secure hardware in benchtop synthesizers with an emphasis on securing the integrity of the screener on a benchtop synthesizer. At the end I provide why even with perfect secure hardware implementation, malicious synthesis is still possible and a difficult problem.

VibeSys: Self-Improving Bespoke Systems
Authors: Keisuke Kamahori, Shihang (Vic) Li, Simon Peter, Baris Kasikci
Presenter: Keisuke Kamahori ↗
Lab/Research Area: SyFI Lab

Abstract: General-purpose systems often fail to fully exploit workload-specific characteristics and heterogeneous hardware, while manual specialization is costly and difficult to scale. We present VibeSys, a multi-agent framework that autonomously builds and continuously optimizes bespoke systems across the computing stack. Its first application, VibeServe, generates LLM serving systems tailored to specific models, hardware, and workloads through coordinated implementation, correctness evaluation, and performance optimization. VibeServe achieves up to 2.3× higher throughput than tuned SGLang on AMD MI300A GPUs and 2.7× higher throughput than the vLLM Neuron plugin on AWS Trainium2. We further demonstrate the potential of AI-driven system specialization in GPU megakernels, microservices, and concurrent data structures. Ongoing work integrates formal verification and simulation to improve reliability and accelerate automated system design.

GlucoScreen-C User Experience Demo
Authors: Jason Hoffman, Kristyna Kalisova, Ananya Vaidyaraman, Sanjana Satagopan, Qiulin Qu, Tammy Nguyen, Kate Muret, Shwetak Patel
Presenter: Kristyna Kalisova ↗, Ananya Vaidyaraman, Sanjana Satagopa ↗n, Jason Hoffman ↗
Lab/Research Area: Ubicomp Lab

Abstract: GlucoScreen-C, as a novel platform for smartphone-based test strips for biomarker screening, moves minimal hardware from the reader into the test strip device, allowing it to communicate the result of electrochemical biomarker detection reactions to the phone without incurring high cost or requiring extra hardware (beyond the test strip itself). This is the first of its kind, and starts to enable a new vision for healthcare where the smartphone becomes the center of testing.

With its common sensors, large touchscreen user interface, and familiar interaction patterns, this can become a much more approachable form of Rapid Diagnostic Testing (RDT) in the future. We invited the CHI community in 2026 to try our prototype demo and share their thoughts and ideas about the future of remote healthcare.

Making Food Access Visible Smart Sensor Monitoring of Grassroots Micro Pantries
Authors: Marianna Rivera, Jose Rivera, Stella Plavsic, Stefany Cruz, Vikram Iyer
Presenter: Marianna Rivera ↗, Stella Plavsic ↗
Lab/Research Area: Vikram Iyer – Micro Computers

Abstract: Community micro pantries have emerged as a grassroots approach to addressing food insecurity and reducing food waste in Seattle. However, the volunteers and organizations maintaining these sites often lack information about food availability, usage patterns, and storage conditions. Without this information, restocking decisions rely on periodic in-person check ins, making it difficult to respond to changing community needs. In this work, we present the design and deployment of a replicable sensing system for monitoring community micro pantries. Our system integrates weight, magnetic door, and temperature sensors to track changes in food availability, pantry access, and refrigerated storage conditions. We deployed 6 pantries across King County to investigate how sensing infrastructure can support more informed food management decisions and the long-term operation of community micro pantries. With 10,644 total pantry activity logs, we cross-compare each pantries performance to best understand how to make each technology probe more resilient against realistic environmental factors faced during the testing period of February to October 2026. We also draw parallels between community micro pantries and other similar community infrastructure to better understand how varying neighborhood socioeconomic factors may improve or hinder the operations of these initiatives.

Extubation Failure Detection Using Acoustic Biomarkers
Authors: Alex Ching, Shwetak Patel, Sarah Wahlster, James Town
Presenter: Alex Ching ↗
Lab/Research Area: Ubicomp Lab

Abstract: Patients who have been intibated due to trauma will often need to be reintubated due to a variety of factors. We are exploring the detection of this extubation failure by utilizing a custom audio recording pipeline and developing a ML pipeline to extract features that can be used for a classification model. We are currently running a study with our partners at Harborview Medical Center for this pilot study.

Social Harness
Authors: Tapan Chugh, Reshabh K. Sharma, Arvind Krishnamurthy, Ratul Mahajan
Presenter: Tapan Chugh ↗
Lab/Research Area: Systems / Agent Infrastructure

Abstract: The Internet will soon have trillions of agents interacting across trust boundaries. Making these interactions efficient and secure requires new infrastructure. Our research reveals a critical gap: today agents can reason and use tools, but they struggle to represent their owners’ interests and defend against manipulation when interacting with other agents. We are building a “social harness” that enables agents to communicate, negotiate, and transact efficiently and safely.

Tilo: Learned Tree Search and Reactive Control for Cluttered Scene Non-Prehensile Manipulation
Authors: Paolo Torrado, Jack Lowry, Joshua Levin, Joshua Migdal and Joshua Smith
Presenter: Paolo Torrado ↗
Lab/Research Area: Sensor Systems Lab

Abstract: Robotic manipulation in cluttered scenes with tree search to warm start model predictive control.

Sensing On A Budget
Authors: Chun-Cheng Chang, Shwetak Patel, Vikram Iyer
Presenter: Chun-Cheng Chang ↗
Lab/Research Area: Ubicomp Lab, Ubiquitous Computing and On-device ML

Abstract: Wearables must recognize what their users are doing while staying within a strict power budget. Common schedulers either ignore that budget or spend it on a fixed clock. Confidence thresholds raise sensing when the classifier is unsure, but stayed within budget in only 27.8% of our runs. We treat the budget as a hard constraint instead. A scheduler trained with PPO-Lagrangian decides, every few seconds, whether each sensor runs at full rate, in a low-power mode that samples less, or not at all. Across five wearable datasets, it meets the budget in 94.4% of runs and reaches tight budgets that duty cycling cannot. We also find that sensing less can make a classifier more accurate on new people. On WESAD, each person’s best sensor setting raises accuracy from 0.56 to 0.76, but that setting differs from person to person. Finding it without that person’s labels remains an open problem.

AugPrint: 3D Printing Foam Augmentations Directly Onto Objects Using Viscous Thread Printing
Authors: Jerry Cao, Yuecheng Peng, Hongrui Wu, Yuxuan Miao, Sanjana Satagopan, Runxin Shi, Brier Hehmeyer, Brett Emery, Jeffrey Lipton, Richard Li, Jennifer Mankoff, and Shwetak Patel
Presenter: Jerry Cao ↗
Lab/Research Area: UbiComp Lab and Make4all Lab (Fabrication, Accessibility, HCI)

Abstract: Combining soft-structure fabrication, direct-to-object augmentation, and embedded interactivity remains challenging, typically requiring specialized hardware, precise 3D scanning, or complex post-processing. AugPrint demonstrates how viscous thread printing (VTP) can bridge all three domains using an unmodified consumer-grade 3D printer. VTP leverages liquid rope coiling to deposit material from a high clearance, thereby inherently mitigating print-head collisions on curved surfaces. The extruded filament remains viscous enough to passively conform to underlying geometries without requiring precise mesh alignment. We present a design tool that enables users to create gradient-density foam augmentations on curved surfaces with non-planar toolpath generation, stiffness control, and surface texturing. We also establish the viability of embedded pressure sensing using commercially available conductive TPU printed with VTP. We demonstrate these capabilities through end-to-end applications, including a pressure-mapping bike seat and a shock-absorbing microcontroller enclosure.

BonVoyage: Better Rewards without Ranking
Authors: Gonçalo Faria, Guang Yang, Alisa Liu, Noah A. Smith
Presenter: Gonçalo Faria ↗
Lab/Research Area: NLP

Abstract: Reward models (RMs) play a central role in reinforcement learning from human
feedback (RLHF), yet they are typically trained with a Bradley–Terry (BT) objective that ranks response pairs under a fixed text distribution. This objective does not match how RMs are used in RLHF: as training progresses, the policy increasingly concentrates on responses that receive high reward under the RM itself. But what matters is not whether the RM ranks fixed pairs correctly, but whether it remains reliable on the responses that its own signal makes more likely under the policy. We develop BONVOYAGE, a method for training RMs directly for this downstream role. Instead of training a pairwise ranker, BONVOYAGE maximizes the likelihood that the RM-induced RLHF-optimal policy assigns to preferred responses. Estimating the gradient of this objective requires samples from the optimal policy induced by the learning RM. We obtain these samples through test-time alignment: given the base LM and the RM, we use Markov chain Monte Carlo to sample from the induced policy. To make this practical, we use the base LM as an independent proposal distribution, reuse candidate responses across optimization steps, and use contrastive divergence with short chains initialized at preferred responses. We evaluate BONVOYAGE by training 4B RMs for QWEN3-1.7B-BASE in mathematics and our supervised finetuned model QWEN3-1.7B-SCIENCE in science. Compared with BT baselines, BONVOYAGE yields stronger downstream policies under RLHF and stronger rerankers under best-of-n sampling across MATH-500, GSM8K, SciQ, SciKnowEval, ARC-Challenge, and MMLU-Science. On verifiable math tasks, BONVOYAGE also shows substantially more robustness to reward over-optimization, preserving alignment between RM scores and ground-truth correctness later into RLHF training

FlexiCodes: Embedded De Bruijn Patterns for Deformation Tracking on Elastic 3D-Printed Objects with a Monocular Camera
Authors: Daniel Campos Zamora, Ria Agarwal*, Raiden Santos*, Jerry Cao, Jon E. Froehlich, Chang Xiao
Presenter: Daniel Campos Zamora ↗
Lab/Research Area: Makeability Lab (HCI)

Abstract: Non-rigid interfaces allow for soft, deformable inputs using malleable materials, which enable interactions that would be infeasible with traditional rigid controllers. However, tracking and decoding the deformations typically requires embedded sensing hardware or complex external setups (e.g., stereo cameras, calibrated projector setups). We present FlexiCodes a novel approach to deformable interaction that leverages 3D printing to embed De Bruijn patterns onto elastic objects and computer vision to track surface deformation.

Orthogonal Trust: The Independence of Privacy and Accuracy in Instructor Evaluations of AI Tools
Authors: Evolone Layne, Nya Domkam, Afi Tagnedji, Shukra Nankabirwa, Azhar Kimanje, Vicente Arroyos, Teanna Barrett, Semayat Yewondwossen, Marquiese Garrett, Brianna Smith, Aisha Cora, Vikram Iyer and Kyle Johnson
Presenter: Marquiese Garrett ↗
Lab/Research Area: Iyer Lab

Abstract: Educational institutions face a tension between generative AI’s affordances and the compliance risks of sending student data to third-party cloud APIs. On-device small language models (SLMs) offer an alternative, guaranteeing data sovereignty through local execution, yet pedagogical viability remains unestablished. We present TOOL, an on-device educational assistant that runs a 1.5-billion-parameter model locally without cloud connectivity, with no-code fine-tuning on instructor materials and built-in Socratic guardrails. Automated benchmarks show the educator-evaluated model achieved 12.4 times higher guidance ratings than its base model while preserving aggregate MMLU performance. To evaluate deployment readiness, we conducted a qualitative case study with seven educators spanning computing, engineering, and education-policy contexts who integrated their own course materials and evaluated TOOL through think-aloud interaction, interviews, and a 30-item survey. Our findings reveal orthogonal trust: data privacy and epistemic accuracy function as independent, non-fungible evaluation axes.

Educators decisively endorsed the local privacy architecture (M = 4.07), but this did not transfer to confidence in factual reliability (M = 2.74); parametric leakage and fluent-but-wrong output led six of seven educators to reject unverified adoption, and five of seven identified traceable citations as non-negotiable for classroom use. While privacy is necessary for institutional consideration, epistemic reliability remains the binding constraint on adoption.

VideoNet: Action Recognition in the VLM Era
Authors: Tanush Yadav, Mohammadreza Salehi, Jae Sung Park, Vivek Ramanujan, Hannaneh Hajishirzi, Yejin Choi, Ali Farhadi, Rohun Tripathi, Ranjay Krishna
Presenter: Tanush Yadav ↗
Lab/Research Area: RAIVN Lab

Abstract: Actions, which transcend multiple frames, are the defining characteristic of video as a medium. Accordingly, for many decades action recognition was the quintessential video understanding task, but recently it has become obsolete. In this work, we revitalize action recognition for the era of VLMs by focusing on fine-grained, domain-specific actions. Frontier models struggle on our benchmark relative to non-expert humans, especially when given few-shot examples.

Tracking What Matters: A Goal-Directed Approach to Health Self-Tracking
Authors: Sean Munson, James Fogarty, Julie Kientz, Shaan Chopra, Yasaman Sefidgar, Tae Jones, Carla Castillo, Katherine Juarez, Jessica Schroeder, Chia-Fang Chung, Daniel Epstein, Ravi Karkar, Natalia Murinova, Allison Cole, Jasmine Zia, Laura Pina, Hyeyoung Ryu, Roger Vilardaga, Qiaosi Wang, Anant Mittal, Liwei Jiang
Presenter: Shaan Chopra ↗
Lab/Research Area: Human Computer Interaction, Health Informatics

Abstract: Nearly 70% of US adults track at least one health indicator, yet most struggle to find value in their data. Current self-tracking tools are built around assumed goals rather than the goals people actually have — a design failure that undermines tracking, behavior change, and patient-provider collaboration. Tools operationalize goals too narrowly, support only a slice of what people need, and do not adapt as goals evolve. To address this, we propose a goal-directed self-tracking process in which clinicians, designers, and developers integrate medical and self-tracking expertise in design to support people in interactively configuring individualized self-tracking. Our work demonstrated this framework across two chronic condition case studies: IBS food trigger identification (via Foodprint for hypothesis formation and TummyTrials for hypothesis testing) and migraine management (via MigraineTracker, developed through surveys of 279 patients, clinical interviews, and a 12+ month deployment with 10 patients and 3 clinicians). We found that goal-directed tracking enabled better self-knowledge, clearer clinical communication, personalized action, and patient confidence in their lived experience.

In recent research, we examine whether generative AI can scale this approach across conditions, populations, and the realities of comorbidity. We conducted preliminary investigations of how existing generative AI tools (e.g., ChatGPT, Copilot) can provide in-the-moment support for goal-directed self-tracking. This work found generative AI to have good potential for supporting people in planning for tracking, in defining tracking goals, and in generating personalized tracking plans, but that people need more scaffolding to reliably benefit. Building on these results, our current research agenda spans goal elicitation support via generative AI conversation, goal-aware prompting for personal health data interpretation, AI-generated clinical summaries aligned with patient goals, and equity-centered design for populations not currently well-served by self-tracking tools.

Change Tours: Code Review via Structured Narratives
Authors: Hannah Potter, Tijana Minic, Andrew Blinn, David Anthony K. Fields, James Yoo, Michael D. Ernst, René Just, Cyrus Omar
Presenter: Hannah Potter ↗
Lab/Research Area: PLSE (Programming Languages and Software Engineering)

Abstract: With the increasing use of generative AI for coding tasks, developers are focusing more on code review to critically evaluate changes and maintain an understanding of their software. However, navigating and comprehending the code changes in a pull request (PR) is challenging. Today’s PR review tools, such as GitHub’s PR browser interface, primarily display changes as a set of file-level diffs, with the changed files listed in directory-alphabetical order. Large changes are difficult to comprehend when presented this way, so authors write a PR description to aid review.

PR descriptions often refer to implementation details and give a high-level guide to what changes were made and sometimes how to review them, but individual code changes are rarely explicitly referenced. This indicates an interface gap in which PR reviewers must themselves transform the narrative provided by the PR description into a navigation plan through the low-level code changes. To address this gap, this paper introduces change tours. A change tour reorganizes a pull request into a hierarchical narrative that interleaves references to every individual hunk (group of nearby changed lines) within the change. This interleaving gives reviewers a clearly ordered navigation plan through the code changes, with each step contextualized in a broader narrative that can be read at varying levels of detail. We built ChangeTours, a system for GitHub PR review implemented as a Visual Studio Code extension, and studied its impact on both PR authors and reviewers in 11 real pull requests in two open-source projects. Participants reported that change tours made reviews easier to navigate and more thorough, with the largest benefits on large pull requests, and that AI assistance made authoring a tour manageable. The comments that reviewers left did not measurably change, which suggests that the effectiveness of a change tour depends on the underlying change and on the quality of the narrative as much as on its presentation.

Event-Driven Low-Power Post-Crash Sensing Network
Authors: Chun-Cheng Chang, Jiayi Shao, Kamran Ali, Jinzhu Chen, Fan Bai, Vikram Iyer
Presenter: Chun-Cheng Chang ↗, Jiayi Shao ↗
Lab/Research Area: Ubicomp Lab, Ubiquitous Computing and Sensing

Abstract: After a crash, first responders often arrive knowing little about what happened inside the vehicle. In collaboration with General Motors, we built a low-power, multimodal post-crash sensing system: wireless sensing nodes distributed across the car stay in a microamp-level sleep while parked, then wake on impact to measure crash severity and direction, count occupants with an on-device camera, and listen for calls for help. Within 1.7 s of an impact, a compact crash report reaches a remote receiver over long-range LoRa radio, and RF wireless charging keeps the nodes powered while the car is parked.

Interactive Multi-modal Agents
Authors: Zixian Ma, Chenhao Zheng, Weikai Huang, Jieyu Zhang, Ranjay Krishna
Presenter: Zixian Ma ↗, Chenhao Zheng ↗
Lab/Research Area: RAIVN / Computer Vision

Abstract: Visual agents face a shared challenge across digital tasks: their actions change what they observe next, and their guidance changes what users do next. We present two complementary steps toward studying that interaction. MolmoWeb develops open visual browser agents that act from screenshots and task instructions, together with data and methods for training and evaluating them. Interactive Visual Gym broadens the evaluation setting to executable visual environments and simulated users, where an assistant’s guidance affects how a task unfolds. Together, these projects chart a path toward multimodal agents that can act reliably in visual environments and collaborate effectively with people.

Reliable and Reconfigurable Molecular Circuits
Authors: Lancelot Wathieu, Zoë Derauf, Chris Thachuk
Presenter: Lancelot Wathieu ↗
Lab/Research Area: Molecular Programming

Abstract: Molecular circuits can sense their environment, perform computation, and trigger biochemical responses, but designing reliable, scalable systems remains challenging. My research combines logic synthesis with the development of leakless DNA strand-displacement architectures, enabling optimized circuits built from reusable molecular components. Through theory, compilers, and wet-lab experiments, I demonstrate molecular circuits that are more reliable and programmable.

ReHearsal: Agentic Drum Performance Reconstruction from Audio and Video
Authors: Teerapat Jenrungrot, Ira Kemelmacher-Shlizerman, Steven M. Seitz
Presenter: Teerapat Jenrungrot ↗
Lab/Research Area: GRAIL

Abstract: ReHearsal reconstructs a drum performance from ordinary audio and video: which drum or cymbal was struck, when, with which hand, and how the sticks and the drummer’s body moved in 3D. The sound decides what was struck. AI agents click a few points on the sticks and the drum kit, and an optimization fits the kit, two rigid sticks and a mannequin to every camera. Naming the toms and cymbals by their sound beats the agents’ visual labels on all four held-out test sets (+.110 to +.184 F1). On a 15-second lab clip, the 3D sticks give the correct hand for 94.5% of hits, against 72.5% for a model that guesses from the hits alone. A prototype agent also builds the drum kit itself as a program, from drums, cymbals and stands that must pass physical and visual checks before the kit is accepted.

How background noise affects small voice AI
Authors: Chun-Cheng Chang, Vidya Srinivas, Zachary Englhardt, Vikram Iyer, Shwetak Patel
Presenter: Chun-Cheng Chang ↗
Lab/Research Area: Ubicomp Lab

Abstract: Voice assistants are moving onto phones, earbuds and glasses, where they must run on small models and work in noisy places. There are two ways to build one: transcribe the speech first and let a language model work from the text, or let one model listen to the audio directly. We compare both designs under controlled background noise, using the same phone-size model (Gemma 3n) in each, and ask whether failures come from mis-hearing (the ear) or mis-reasoning (the brain). For this model, transcribing first wins on all six tasks at every noise level: on clean audio, it solves 93% of short spoken math questions from a transcript but only 15% when it listens itself. For spoken math, the failures trace back to the ear. When every number is transcribed correctly, answers nearly match a perfect transcript (95% vs. 98%); when even one number is misheard, fewer than 1% are correct. Across the systems we tested, other people talking was the most damaging noise.

Size Conditioned Virtual Try-On
Authors: Yuanhao Wang, Varsha Ramakrishnan, Steve Seitz, Ira Kemelmacher-Shlizerman
Presenter: Yuanhao Wang ↗
Lab/Research Area: UW Reality Lab/Computer Vision

Abstract: Real-world clothing try-on serves two purposes: visualizing how a garment looks on the body, and seeing how a specific size actually fits—whether it stretches tightly or drapes loosely. Current Virtual Try-On (VTO) models address appearance but ignore fit; conditioned only on 2D images, they default to synthesizing deceptively well-fitted results regardless of garment dimensions. In this work, we present RealFIT, the first scalable framework for real-world, measurement-conditioned virtual try-on across complex apparel categories. To overcome the scarcity of metric-labeled data and the sim-to-real gap of 3D physics engines, we introduce a multi-tier data hierarchy combining web retail metadata, VLM-assisted annotations, and a newly curated real-world dataset with tape-measured ground truth. We further propose a flow-matching VTO architecture featuring a dedicated measurement encoder and a Localized Metric Contrastive (LMC) loss that contrasts velocity fields across sizing perturbations to sharpen fit sensitivity. Finally, we establish a systematic evaluation protocol and demonstrate that RealFIT achieves state-of-the-art photorealism, fit accuracy, and size controllability.

AutoStringer
Authors: Shritan Gopu, Jason Le, Pranav Sharma, and Justin Wu
Presenter: Shritan Gopu ↗ and Jason Le ↗
Lab/Research Area: Robotics

Abstract: Fully autonomous racket stringing machine