Allen School Ph.D. student Shangbin Feng aims to build a more open and democratic artificial intelligence future. To that end, his research focuses on model collaboration, where “multiple AI models, trained on different data, by different people, and thus possess diverse skills and strengths, collaborate, compose and complement each other.”
In December, Feng was named among the 2026 class of NVIDIA Graduate Fellows in recognition of his work. The NVIDIA Graduate Fellowship program supports graduate students from around the world whose outstanding research puts them at the forefront of accelerated computing and is relevant to the company’s interests.
“Through model collaboration, I aim to spearhead a modular, compositional, decentralized, and participatory AI future — that everyone everywhere could have a say in the future of AI by contributing data, models, or natural language feedback reflecting their interests and priorities, and building a compositional AI system from the bottom up with their decentralized contributions,” said Feng, who is advised by Allen School professor Yulia Tsvetkov.
Feng used model collaboration techniques to enhance the reliability and trustworthiness of large language models (LLMs). Despite evolving efforts to expand the LLMs’ knowledge base, they still run into knowledge gaps, or missing or outdated information. To help models abstain from generating low-confidence outputs, he and his team introduced two novel, robust multi-LLM collaboration-based approaches where LLMs probe other LLMs for knowledge gaps, either cooperatively or competitively. When multiple LLMs are working together in cooperation, one LLM employs other models to give feedback on the proposed answer, and it then synthesizes all the outputs into an overall abstain decision. In a competitive setting, the LLM is challenged by other models with conflicting information, and it has to decide whether to abstain or not. The team’s paper describing this approach received an Outstanding Paper Award at ACL 2024, the conference organized by the Association for Computational Linguistics.
In the same vein of competitive LLM collaboration, Feng helped introduce Sparta Alignment, a framework that collectively aligns multiple language models through combat and game theory. Models form a “sparta tribe” to battle against, evaluate and learn from the strengths and weaknesses of each other. In each iteration, a pair of models duel by generating responses to sampled prompts from the dataset, while the remaining models judge their outputs. As models win or lose battles their reputation shifts, impacting how much say they have in evaluating other LLMs. For Feng, Sparta Alignment “enables the collaborative evolution of diverse LLMs without external supervision.”
He has also utilized multi-LLM collaboration to help pluralistically align models to better reflect the diversity of human values, intentions and preferences. Additionally, Feng developed the collaborative search algorithm called Model Swarms, in which diverse LLM experts collectively move in the parameter search space using swarm intelligence.
“Shangbin works on ‘model collaboration,’ a research program he is pioneering,” said Tsvetkov. “Advancing this program, Shangbin has already achieved a highly prolific publication record with peer-reviewed papers (mostly first-author) in top conferences.”
That record includes the aforementioned Outstanding Paper Award at ACL 2024, the top conference in natural language processing; a Best Paper Award at ACL 2023; a spotlight paper at the 37th Annual Conference on Neural Information Processing Systems (NeurIPS 2023), the top machine learning conference; an oral presentation at the 12th International Conference on Learning Representations (ICLR 2024), representing the top 1.2% of accepted papers; and an Area Chair Award in the QA Track at ACL 2024 (one of the most competitive track in NLP conferences).
At the recent ACM CHI Conference on Human Factors in Computing Systems (CHI 2026), Allen School researchers brought home multiple accolades for their innovative work in human-computer interaction (HCI) and artificial intelligence. Their projects ranged from interactive systems that allow users to collaborate with AI agents with more flexibility, to an AI-based tool that helps screen-reader users make sense of geovisualizations, to a method for customizing LLM outputs based on user objectives — and much more.
Best Paper Award: Cocoa
As AI agents take on more complex tasks that require sophisticated planning and execution, such as writing research reviews or analyzing complex documents, there is a need for more users to work together with AI to tackle these problems. However, existing agentic research tools only focus on supporting human-AI collaboration either before or after task execution.
A team of researchers including Allen School professor Amy X. Zhang introduced Cocoa, an interactive system that enables scientific researchers to co-plan and co-execute alongside AI agents in a document editor to tackle open questions and tasks within their research projects. With Cocoa, users and AI agents can jointly complete plan steps and then re-execute steps as desired — similar to executing code cells in a computational notebook.
Amy X. Zhang
“In this work, we chose to really emphasize flexibility in designing an interface for working with a long-running AI agent so as to give users more control — they can switch from planning to execution and back again, and can also take over from the agent to alter plans or assist in execution at any step along the way,” said Zhang.
Based on a formative study about the needs of researchers who use AI to support their work, the team designed a user interface that provides flexible delegation between human and AI work. For example, in the interactive sidebar, users can edit the AI agent’s outputs and add any relevant papers that the agent did not find to help guide its expertise and feedback. The user can also use the step assignment toggle feature to assign low risk, but high effort tasks to the AI agent, such as searching for papers or expanding on preliminary ideas. Then, the user can reserve the more consequential tasks that require higher thinking for themselves, including identifying seed papers for the AI agent to explore more broadly and making connections between multiple papers.
The team first evaluated Cocoa against a custom chat baseline in a within-subjects task-based lab study with 16 researchers. Participants noted that the system enabled greater steerability without sacrificing ease-of-use compared to the strong custom chat baseline. In a week-long field deployment study, seven participants integrated cocoa in their day-to-day research and found that the system was especially helpful for literature discovery and synthesis,
Additional authors include Allen School professor emeritus Daniel Weld; University of Washington Human Centered Design & Engineering Ph.D. student Kevin Feng; University of Toronto Ph.D. student Kevin Pu; Matt Latzke, Pao Siangliulue, Jonathan Bragg and Joseph Chee Chang at the Allen Institute for Artificial Intelligence (Ai2); and Tal August, faculty at the University of Illinois Urbana-Champaign.
Geovisualizations, or interactive map visualizations, are powerful tools for understanding patterns and trends in spatial data, but they are inaccessible to screen-reader users. Even when accessibility features such as alt text and data tables are available, these features struggle to capture and interpret the full potential of complex geovisualizations.
To address this gap, a team of researchers in the Allen School’s Makeability Lab introduced GeoVisA11y, an AI-based question-answering system that makes geovisualizations more accessible using natural language interaction. The system integrates geostatistical analysis with large language models (LLMs) to go beyond keyword-matching and rule-based approaches and make previously inaccessible analytical tasks possible.
Chu Li
“We started by exploring how screen-reader users could access interactive maps, but what we found is that natural language interaction with geospatial data benefits everyone. Accessibility is never binary, and we should create tools that support people across a whole spectrum of spatial analysis abilities,” said Allen School Ph.D. student and lead author Chu Li, who is advised by Allen School professor and senior author Jon Froehlich.
GeoVisA11y is made up of two primary components. The first combines a screenreader-compatible user interface that includes an interactive map with an AI-based chat tool that can answer analytical, geospatial, visual and contextual questions. The team paired that with a custom question-answering pipeline that can transform natural language questions into geoanalytical responses and map interactions. For example, if a user asks if there is a pattern on the map, the system first runs a global Moran’s I test, which can detect significant patterns using spatial autocorrelation, and if autocorrelations exist, it then performs a local indicators of spatial association (LISA) analysis to identify specific spatial clusters or outliers. These LISA outputs are then summarized by GPT and presented to the user along with representative examples.
The researchers evaluated GeoVisA11y through a series of user studies in which six screen-reader users and six sighted participants were asked to complete various data analytics tasks. Both groups engaged with the geospatial data through GeoVisA11y with different, but complementary, strategies. The screen-reader users employed a combination of verbal queries and keyboard navigation, while the sighted users visually assessed the map first, then queried for specific details. Despite their varying approaches, both groups completed the tasks and identified similar patterns, showing that GeoVisA11y could effectively bridge accessibility gaps and create a shared understanding of geovisualizations.
In addition to the two Best Paper Awards, Allen School authors received four honorable mentions at CHI 2026 for their research.
A team of researchers in the Mobile Intelligence Lab led by Allen School professor Shyam Gollakota were recognized for VueBuds, the first system that incorporates small cameras into off-the-shelf wireless earbuds to allow users to chat with an AI model about the scene in front of them. For privacy, the prototype only takes still images that are stored and processed on the device, and which the user can delete immediately. On the topic of prototypes, user studies in HCI often evaluate whether a prototype is “better,” however, the perceived newness of technologies can influence users’ judgement and possibly performance. Allen School Ph.D. student Yumeng Ma and her collaborators were recognized for their paper quantifying this novelty bias.
Heer along with colleagues at Stanford University received an honorable mention for introducing just-in-time objectives, a new method for automatically inducing AI objectives on the fly based on observing the user and their task. This approach allows the LLM to produce customized rather than generic outputs, from individual responses to generated software tools. Meanwhile, as conversational LLM interfaces become more commonplace in data analysis, the challenge becomes how can data workers easily go back, make sense of long analytical conversations and then communicate their insights to others. Allen School Ph.D. student Ken Gu and collaborators at Tableau Research were recognized for their paper investigating how data workers revisit these conversations and what kinds of tools can support that process.
Also at the conference, Allen School professor and alum Jon Froehlich (Ph.D., ‘11) collected a SIGCHI Societal Impact Award for his work tackling accessibility challenges through HCI and AI, while fellow alum Jeffrey Bigham (Ph.D., ‘09) was inducted into the SIGCHI Academy.
As the director of the University of Washington’s Makeability Lab, Allen School professor and alum Jon Froehlich (Ph.D., ‘11) utilizes human-computer interaction (HCI) and machine learning to tackle high-impact socially relevant problems. Already, his work has led to improved city planning and sidewalk infrastructure across the globe, and he has developed technologies that have enabled blind and low-vision users to prepare meals, participate in sports and even engage with children’s artwork.
“Our work aims to transform how humans interact in the real world via advanced techniques in HCI and AI, such as assessing the bikeability of a city at scale using vision language models and providing personalized bike routes, determining whether a building is accessible and to whom via capability-conditioned AI agents, and allowing people who are deaf or hard of hearing customize their own sound feedback visualizations with Generative AI,” Froehlich said. “This is an incredible time to work in HCI because sensing hardware, computation, and processing have transformed how we can augment human capabilities in the world.”
The ACM Special Interest Group on Computer-Human Interaction (SIGCHI) recently honored Froehlich with the 2026 SIGCHI Societal Impact Award, which recognizes mid-career to senior researchers whose HCI work “demonstrates social benefit.”
“This recognition belongs to my incredible students and collaborators in the Makeability Lab who work tirelessly to design more accessible, equitable futures and pursue research in accessibility, education, and environmental sustainability,” said Froehlich, who is also an associate director of CREATE (Center for Research and Education on Accessible Technology and Experiences).
At the heart of Froehlich’s societal impact is his work developing Project Sidewalk, a web-based platform that uses crowdsourcing and AI to transform how sidewalks are visualized, mapped and analyzed. Users can explore cities through immersive imagery, similar to a first-person video game, and label sidewalks as well as accessibility issues such as uneven surfaces, missing curb ramps or obstructions like poles or fire hydrants. Project Sidewalk has assembled the largest ever sidewalk accessibility dataset — it includes more than 3 million data points across 27,000 kilometers of city streets in 43 cities and 10 countries. Froehlich and his collaborators presented a paper on the project’s pilot deployment in Washington D.C. at CHI 2019, where they received a Best Paper Award.
Across the country, cities have used Project Sidewalk to help fund initiatives to improve the safety and accessibility of their pedestrian infrastructure. For example, in Newberg, Oregon, Project Sidewalk collected more than 17,000 labels and showed that sidewalks were inaccessible especially around voting centers and bus stops, prompting the city council to authorize $50,000 for immediate sidewalk repairs and establish a grant program to help homeowners fix their own sidewalks. After Mendota, Illinois, experienced devastating fires in 2022, community partners used Project Sidewalk data to secure a $3.6 million Illinois Transportation Enhancement Program grant to rebuild their sidewalks.
Project Sidewalk has also helped spur systemic changes in how governments make decisions about accessibility in their communities. In Chicago, the Project Sidewalk team provided expert testimony that helped shift the city’s infrastructure allocation from a complaint-based system to a more data-driven prioritization. The Mexico-based NGO Liga Peatonal partnered with Project Sidewalk to connect municipal accessibility data to international development frameworks. Froehlich also leveraged Project Sidewalk data in his collaboration with the University of Zurich in Switzerland on the ZüriACT project, which aimed to make Zurich more accessible and walkable.
Outside of Project Sidewalk, Froehlich has worked on a variety of other accessibility projects. He led the development of StreetReaderAI, the first screen reader for Google Street View which can describe the physical environment such as sidewalks, crosswalks and ramps to blind and low-vision users. Froehlich and his collaborators also introduced CookAR, a headmounted augmented reality system that provides low-vision users real-time support during meal preparation, and earned the Belonging and Inclusion Best Paper Award at the ACM Symposium on User Interface Software and Technology (UIST 2024). More recently, he and his team developed GeoVisA11y, an AI-based geovisualization question-answering system for screenreader users, which received a CHI 2026 Best Paper Award.
“Jon Froehlich’s research in urban accessibility has achieved what few HCI researchers ever accomplish: direct, measurable change on a global scale,” said Jacob Wobbrock, UW Information School professor and Allen School adjunct faculty. “Specifically, his work has affected how cities invest in pedestrian infrastructure, how communities and governments in 40+ cities plan for accessibility, and how federal agencies define walkability and accessibility data standards.”
Froehlich joins a long line of recipients of the SIGCHI Societal Impact Award from the Allen School community. This includes faculty colleague Jennifer Mankoff, professor emeritus Richard Ladner, alum Nicki Dell (Ph.D., ‘15) and Wobbrock. He will be formally honored at the CHI 2026 conference taking place in Barcelona, Spain, later this month.
Also at the conference, fellow Allen School alum Jeffrey Bigham (Ph.D., ‘09) will be inducted into the SIGCHI Academy. Bigham, who is now faculty at Carnegie Mellon University and director of human-centered intelligence and responsible AI at Apple, is being recognized for his research that combines machine learning and crowdsourcing to enable people to interact with systems in a useful and responsible way, with a focus on accessibility.
Allen School professor emeritus Richard Ladner holds the SIGCSE Award for Outstanding Contribution to Computer Science Education next to Paul Tymann, faculty at the Rochester Institute of Technology, who received the SIGCSE Award for Distinguished Service to the Computer Science Education Community.
Although Allen School professor emeritus Richard Ladner started his career as one of the founders of the University of Washington’s theoretical computer science group, he has grown to become a leading researcher in accessible technology and an advocate for expanding access to computer science for students with disabilities. As a child of deaf educators, Ladner is deeply committed to accessibility and disability inclusion, and credits his first-hand experience with making him a more effective researcher.
“Computing technology has changed the world and almost everyone now has a computer in their pocket,” said Ladner. “It is important that future engineers and technologists learn about accessibility so that they can design technology that is accessible so that everyone can use it, and so that they can build things that are particularly useful to people with disabilities such as screen readers.”
Through mentoring and advocacy, Ladner has also directly helped hundreds of students with disabilities to gain exposure to computing as a potential career path. The Association for Computing Machinery’s Special Interest Group on Computer Science Education (SIGCSE) recently recognized Ladner with the ACM SIGCSE Award for Outstanding Contribution to Computer Science Education. The award honors researchers who have made a long-lasting impact and significant contribution to computing education.
From 2006 to 2024, Ladner led the Alliance for Access to Computing Careers (AccessComputing) — which he co-founded with Sheryl Burghstahler, former director of the UW’s DO-IT Center, with funding from National Science Foundation’s Broadening Participation in Computing program. AccessComputing supports high school, undergraduate and graduate students to build skills and connections with mentors and professional opportunities in computing-related fields. Between 2015 to 2024, program mentors worked with over 1,500 students with disabilities nationwide.
During this same time period, Ladner also led the Summer Academy for Advancing Deaf and Hard of Hearing in Computing. Students enrolled in the program, which ran from 2007 to 2013, jumpstarted their academic careers by spending the summer taking computing courses at the UW’s Seattle campus. Many of the students who went through the summer academy went on to earn bachelor’s and graduate degrees in computer science — including three who became computing faculty themselves. Ladner also partnered with Andreas Stefik, a professor at the University of Nevada, Las Vegas, to launch AccessCSforAll, an initiative that provides accessible computer science curriculum and other resources available to K–12 students with disabilities. Their paper describing the work received the SIGCSE 2019 Best Paper Award in the Experience Reports and Tools Track.
For Ladner, it was not only important that computing education was accessible to students with disabilities, but that teaching about accessibility and disability was included in computer science curricula. In a 2018 SIGCSE paper, Ladner and his collaborators found that around 2.5% of computing and information science faculty nationwide reported teaching about accessibility, with many citing lack of knowledge as a barrier. To help address this gap, in 2024, Ladner co-edited the book “Teaching Accessibility” along with Alannah Oleson, faculty at the University of Denver, and Amy Ko, professor in the UW Information School and adjunct faculty in the Allen School. Covering multiple fields, from robotics to data structures, the book provides computing instructors with the tools and resources to help them incorporate accessibility into their teaching. Ladner has also co-authored several papers on the importance of teaching and learning about accessibility, as well as outlining strategies for integrating accessibility into various computer science courses.
To better understand the impact of these initiatives, Ladner has advocated for the availability of demographic disability data. He encouraged the Computing Research Association (CRA) to report disability data as part of its annual Taulbee Survey, which it has as of 2021. At the same time, he also encouraged the publishers of the State of Computer Science Education report to include data on students with disabilities; they had reported the data annually since 2020. For Ladner, one of the big successes of AccessComputing was getting organizations such as the CRA to “recognize disability as a minoritized group and to collect data about it.”
“Richard is quick to encourage folks who haven’t thought about disability inclusion to find ways to incorporate it in their work — and will have a suggestion for how to do it! Without Richard’s relentless encouragement, accessibility and disability would not be as prominent topics in CS education as they are,” said Brianna Blaser, director of AccessComputing.
Even after his retirement in 2017, Ladner has remained active in accessibility research, mentoring and advocacy. Over the years, he has supervised or co-supervised 30 Ph.D. students as well as more than 100 undergraduate researchers — many of whom sought him out for his expertise in accessibility. One of his students went on to establish the Richard E. Ladner Endowed Professorship, currently held by his faculty colleague Jennifer Mankoff, in his honor. His former students also initiated the Richard Ladner Endowed Fellowship that supports graduate students working in accessibility research.
“Whether students with disabilities, or committed allies, Richard has supported and empowered others to be disability justice advocates in industry, academia and communities. Much of his mark is in the exponentially growing network of scholars, educators, and advocates that he’s supported for decades,” said Ko.
Er-Cheng Tang (right) and Mi-Ying (Miryam) Huang hold the plaque they received for the Machtey Award for Best Student Paper.
Program obfuscation, which aims to obscure the inner workings of a computer program while maintaining its functionality, is a central goal in cryptography and software protection. An emerging line of research explores the possibility of applying obfuscation to quantum programs. So far, however, researchers have only achieved obfuscation in specific quantum circuits — falling short of supporting obfuscation of general quantum input-output functionalities.
Enter Allen School Ph.D. student Er-Cheng Tang and collaborator Mi-Ying (Miryam) Huang, a Ph.D. student at the University of Southern California. The duo recently developed the first quantum state obfuscation scheme for unitary quantum programs, which are the backbone of quantum computing, in the classical oracle model.
“We achieve program obfuscation in the fully quantum setting for the first time, enabling software that runs on quantum data to be provably protected,” said Tang, who is advised by Allen School professors Andrea Coladangelo and Huijia (Rachel) Lin.
Tang and Huang presented their research at the 66th IEEE Symposium on Foundations of Computer Science (FOCS 2025) last December in Sydney, Australia, where they received the Machtey Award for Best Student Paper.
“Obfuscation of quantum programs secure against quantum adversaries is a significantly more powerful extension of obfuscation in the classical world. It’s a wonderful feat achieved by two Ph.D. students,” said Lin.
Building off of previous frameworks for quantum obfuscation, Tang and Huang offer several improvements in their scheme to extend obfuscation to quantum programs with quantum inputs and outputs. They start by building a strengthened cryptographic tool, called a functional quantum authentication scheme, for protecting quantum programs while enabling their execution. To support quantum programs with quantum inputs and outputs, the researchers also integrate quantum teleportation into the framework, allowing the transitions of input and output quantum states between protected and unprotected forms.
At the core of their obfuscation scheme is a novel compiler which serves two major purposes. First, it represents an arbitrary quantum circuit as a projective linear-plus-measurement quantum program, as the functional quantum authentication scheme natively works under that format. The attained projective property provides a basis for analyzing the program’s execution trace. The researchers then prove that their obfuscated unitary quantum program can only be used to compute the implemented unitary transformation and its inverse; nothing else can be derived from the obfuscated program.
“The significance of the result is that it relates obfuscation of general quantum programs — that can take as input quantum states and output quantum states — to obfuscation of classical programs. At a high level, it says that there is a generic way to bootstrap the latter to obtain the former,” said Coladangelo.
As a computer systems researcher, Allen School professor and alum Ratul Mahajan (Ph.D., ‘05) has helped develop technologies powering the networks that support multiple aspects of modern society — from operating online banking accounts to scrolling social media.
The Association for Computing Machinery (ACM) recognized Mahajan among its 2025 class of ACM Fellows for his groundbreaking “contributions to network verification and network control systems and their transfer to industrial practice.” The ACM Fellows are selected by their peers and represent the top 1% of members who have provided notable technical innovations and/or service to the field of computing.
“I love developing systems with new operational paradigms that can bring about a step change in efficiency or performance and developing techniques that provide strong guarantees about robustness of large-scale systems,” said Mahajan, a member of the Allen School’s Computer Systems Lab and co-director of the UW Center for the Future of Cloud Infrastructure (FOCI). “And I doubly love it when my work has real-world impact and changes practice.”
Mahajan has helped make network verification techniques mainstream across both industry and academia. He introduced Batfish, an open source network configuration analysis tool that proactively ensures that planned network configuration changes operate as intended. His work on Batfish, one of the earliest and most widely-adopted network verification platforms, was recognized with the SIGCOMM Networking Systems Award. In collaboration with colleagues across academia and industry, Mahajan has developed several other control plane analysis technologies. Minesweeper is able to verify the accuracy of all combinations of external routing messages and failures; Bonsai speeds up analysis by leveraging symmetries within large networks; and ARC does that via specialized graph-based encodings.
The technology behind Batfish and his other methods were commercialized by Intentionet, where Mahajan was the co-founder and CEO until the company was later acquired by Amazon. Today, more than 75 companies worldwide rely on Intentionet’s technology to help design and test their networks.
In his other vein of work, Mahajan focuses on developing systems that allow for the direct control of large-scale infrastructure. Historically, the control of large-scale infrastructure has been indirect, requiring engineers to tweak local, low-level parameters for routers to generate global system behavior. However, as more and more cloud operators provide their own global infrastructure, indirect control can lead to poor efficiency and reliability. Before joining the Allen School faculty, Mahajan was at Microsoft, where he and his collaborators developed SWAN. That system increases the utilization of inter-datacenter networks by directly controlling the amount of traffic each service sends, as well as frequently reconfiguring the dataplane to match traffic demands. Prior to SWAN, using indirect control settings, the busiest links had an average utilization of between 40 to 60%; the switch to SWAN achieved an almost 100% utilization rate.
Mahajan has also extended direct control to the entire infrastructure for delivering online services. He and his team introduced the Footprint system, which leverages dynamics of an integrated setting to boost efficiency and performance. In simulations partially deploying Footprint in the Microsoft infrastructure, they found that it could carry at least 50% more traffic and reduce user delays by at least 30% compared to current methods. Mahajan also helped develop Statesman, a network-management service that enables multiple direct-control applications to safely operate over shared infrastructure.
“The centralized management and control of network infrastructure that Ratul developed transformed Microsoft’s cloud infrastructure,” said Victor Bahl, technical fellow and chief technical officer for Azure operations at the company. “SWAN now carries over 90% of the traffic in and out of Microsoft’s datacenters, a footprint spanning over 280,000 kilometers of fiber and over 150 points of presence across all Azure regions; Footprint has evolved into Microsoft’s content distribution network; and Statesman is deployed across most Microsoft datacenters.”
“At the Allen School, we have developed a sequence of seminars that introduce undergraduate students to research and give them the opportunity to work on a hands-on research project with a faculty or graduate student mentor,” said Allen School professor Leilani Battle, who co-chairs the undergraduate research committee alongside colleague Maya Cakmak.
“Through these research opportunities, students see a side of computer science that they may not encounter in their classes or internships such as learning how to identify new research problems or learning to draw connections between different areas of computer science,” Battle continued.
From developing policies to help robots learn to introducing methods to better understand the training data behind large language models (LLMs), these nationally recognized students have demonstrated how to take what they learned in the classroom and make a real-world difference.
Haoquan Fang: Enhancing the reasoning capabilities of robots
Haoquan Fang
Award winner Haoquan Fang aims to tackle one of the significant challenges in today’s embodied AI models — how to equip robots with robust, generalizable and interpretable reasoning abilities. In his work, Fang bridges these domains and paves the way for robots that can understand the world and act with purpose.
“I am broadly interested in robot learning. In particular, I focus on developing generalist robotic manipulation policies that leverage strong priors, by optimizing both the data curation and model architectures,” said Fang.
Fang proposed a new model that integrates memory into robotic manipulation. Alongside Ranjay Krishna, Fang spearheaded the development of SAM2Act, a multi-view robotic transformer-based policy that leverages the visual foundation model Segment Anything Model 2 (SAM2) to achieve state-of-the-art performance on existing benchmarks for robotic manipulation. He then built off that architecture to introduce SAM2Act+, which extends the SAM2Act system with memory-based components. This policy enables the agent to better predict actions based on past spatial information, thus enhancing performance in sequential decision-making tasks. Fang and his collaborators published this work at the 42nd International Conference on Machine Learning (ICML 2025), and received the Best Paper Award at the RemembeRL Workshop at the 9th Annual Conference on Robot Learning (CoRL 2025). Last year, his senior thesis on SAM2Act was awarded Best Senior Thesis Honorable Mention from the Allen School.
Fang also co-led the introduction of the first fully open action reasoning model for robotics with MolmoAct. The model, which was designed by a team of University of Washington and Allen Institute for Artificial Intelligence (Ai2) researchers, enables robots to interpret and understand instructions, sense their environment, generate spatial plans, and then execute them as goal-directed trajectories. Across various benchmarks, MolmoAct outperformed multiple competitive baselines, including NVIDIA’s GR00T N1.5. MolmoAct received the People’s Choice Award at the Allen School’s 2025 Research Showcase, as well as the Best Paper Award runner-up at the Rational Robots Workshop at CoRL 2025. The State of AI Report 2025 also highlighted MolmoAct for setting the standard of embodied reasoning, which was later adopted by Google Gemini Robotics 1.5.
Hao Xu: Making internet-scale corpora searchable
Hao Xu
Large language model behavior is shaped by their training data and tokenization. For Hao Xu, understanding the composition of these models’ training data is increasingly more important as the “data scales beyond what is practical to inspect.” Today’s LLM are trained on massive, Internet-scale text datasets, however, it is difficult to analyze and understand the quality and content of these corpora.
“My research interests lie in natural language processing with a focus on large language models. My future work aims to develop more efficient model-data interactions that move beyond today’s brute-force training paradigm,” said Xu. “As a violinist, I also view music as a distinct form of language and am interested in studying how it can be modeled and learned using language modeling techniques.”
Xu’s primary research focuses on bridging this gap. Alongside Allen School professors Hannaneh Hajishirzi and Noah A. Smith and Ph.D. student Jiacheng Liu, Xu developed infini-gram mini, an efficient exact-match search engine that is designed to work on internet-scale corpora with minimal storage needs. The system makes several open source corpora, such as Common Crawl, searchable, and it currently hosts the largest body of searchable text in the open-source community.
Using infini-gram mini, Xu and her collaborators revealed significant widespread contamination across standard LLM evaluation benchmarks. This is where the LLM training data inadvertently contains the test data. Their results raise concerns on how researchers measure artificial intelligence progress, and have sparked new conversations about evaluation integrity and LLM dataset transparency. As lead author, Xu presented the research at last year’s Conference on Empirical Methods in Natural Language Processing (EMNLP 2025) where she and the team received the Best Paper Award.
Xu is also interested in understanding the fundamentals of LLMs, such as tokenization that “shapes how the models interact with text.” She undertook the first systematic examination of how tokenization mechanisms prevalent in English fail in other languages with different morphology or writing systems. The paper presenting these findings, for which she was first author, is currently under review.
Kaiyuan Liu: Strengthening the reasoning capabilities of LLMs
Kaiyuan Liu
Kaiyuan Liu aims to build reasoning-capable AI models. His background in competitive programming informs his research as he develops tests and benchmarks for LLMs’ proficiency in reasoning and self-correction.
“My research goal is to understand and improve the reasoning capabilities of large language models,” said Liu. “This goal emerges from two converging paths: years of competitive programming, which trained me to value algorithmic precision and creativity, and a broader curiosity about how intelligent systems — biological and artificial — learn, reason and cooperate.”
Writing competitive programming problems is a time-consuming task, requiring programmers to set multiple variables and constraints including input distributions, edge cases and specific algorithm targets. This makes it an ideal test for general LLM capabilities. Liu and his collaborators, including Allen School professor Natasha Jaques, developed AutoCode, a closed-loop multi-role framework that automates the entire process of competitive programming problem creation and evaluation. AutoCode can detect with 91% accuracy whether or not a program is a valid solution to a given algorithmic problem. The framework has potential industry usefulness especially as more and more large companies attempt to leverage LLMs to write and submit code independently. The team will present their findings at the 14th International Conference on Learning Representations (ICLR 2026) in April.
Liu also helped develop a set of benchmarks to better evaluate LLMs’ reasoning capabilities in competitive programming. LiveCodeBench Pro is composed of high-quality, live-sourced programming problems from sources such as Codeforces and the International Collegiate Programming Contest (ICPC) that vary in difficulty and are continuously updated to reduce the chance of data contamination. The researchers paired large-scale LLM studies with expert annotations and found that frontier models are proficient in solving implementation-oriented problems, but struggle with complex algorithmic reasoning, nuanced problem-solving and handling edge cases — failing on some of the benchmark’s most difficult problems. LiveCodeBench Pro has already had industry impact as the benchmark was selected for the Gemini 3 launch evaluation. Liu and his collaborators presented LiveCodeBench Pro at the 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025) last December.
Outside of his research, Liu and his team were ICPC 2024 World Finalists for competitive programming. More recently, he coached the UW programming team Jiakrico that won first place at the ICPC PNW Regional competition last November.
Lindsey Wei: Building efficient LLM-powered data systems
Lindsey Wei
Modern data systems power nearly every aspect of our digital world, yet growing data complexity and heterogeneity make it increasingly difficult for systems to consistently interpret and process data at scale. Lindsey Wei focuses on developing “intelligent and reliable data systems that reason about data semantics to make data-driven decision-making more accessible.”
“Large language models open up new opportunities for how systems understand and interact with data,” said Wei. “But integrating these capabilities into data systems in a systematic way remains challenging.”
One setting where these challenges arise is table understanding, which focuses on recovering missing semantic metadata from web tables and is crucial for data integration. Existing LLM-based methods have limitations including hallucinations and lack of domain-specific knowledge. To address this, Wei, alongside Allen School professor and director Magdalena Balazinska, developed RACOON, a framework that augments LLMs with facts retrieved from a knowledge graph through retrieval-augmented generation (RAG) to significantly improve zero-shot performance. Next, Wei aims to extend the system via RACOON+, further improving its accuracy and robustness by strengthening how models link to and reason over external knowledge.
Inspired by how inference-time techniques such as RAG can unlock LLMs’ reasoning over structured data, Wei began exploring how to extend these reasoning capabilities to unstructured data processing — a longstanding challenge in data management. With a team of University of California, Berkeley and Google researchers, Wei developed MOAR (Multi-Objective Agentic Rewrites), a new optimizer for DocETL, an open-source system for LLM-powered unstructured data processing at scale. MOAR introduces a global search algorithm that explores a vast space of possible pipeline rewrites to identify those with the best accuracy–cost tradeoffs under a limited evaluation budget. In experiments across six real-world workloads, MOAR consistently discovered pipelines that were both more accurate and significantly cheaper than prior approaches. The team recently released a preprint of this work, highlighting the need to rethink how optimization is designed for LLM-powered data systems.
In addition to designing LLM-powered data systems, Wei has also helped develop a graphical user interface for MaskSearch, a system that accelerates queries over databases of machine learning-generated image masks, leading to improved model debugging and analysis workflows.
From left to right: Laura Sanità from Bocconi University, Yang Liu from Carnegie Mellon University, Allen School professor Thomas Rothvoss, along with Alon Rosen and Riccardo Zecchina from Bocconi University.
Allen School professor Thomas Rothvoss has carved a career out of complexity. As a member of the school’s Theory of Computation Group, Rothvoss examines the theoretical limits of computer algorithms that are designed to analyze large, complex datasets. His aim is to settle long-standing problems in the field of combinatorial optimization — an area where he has made notable progress since his arrival at the University of Washington in 2014.
“I work in theoretical computer science and discrete optimization,” said Rothvoss, who also holds the Craig McKibben and Sarah Merner Professorship in the UW Department of Mathematics. “Over the years my focus has changed a bit. During my Ph.D., I worked on approximation algorithms which deal with finding provably good solutions to NP-hard problems in polynomial time. Later, I moved more towards discrepancy theory and theoretical aspects of linear and integer programming.”
Last month, Rothvoss collected the inaugural Trevisan Prize in the mid-career category for his breakthrough contributions in the study of optimization problems. The award, which is sponsored by the Bocconi University Department of Computing Sciences and the Italian Academy of Sciences, is named for the late Luca Trevisan in recognition of his major innovations to computing theory.
Rothvoss’ own list of innovations includes a new approximation algorithm for solving the bin-packing problem in polynomial time, one of computer science’s classical combinatorial optimization problems. In the bin-packing problem, the goal is to find the fewest number of identical bins that can hold a list of items, without exceeding each bin’s capacity. By leveraging a combination of a novel two-stage packing method — where first they pack items into containers and then put those containers into bins — and making use of discrepancy theory techniques, Rothvoss and his former UW Department of Mathematics Ph.D. student and current postdoc Rebecca Hoberg introduced an algorithm that achieves an additive gap of only O(log OPT) bins, significantly improving on previous results. The paper received the Best Paper Award at the 25th annual ACM-SIAM Symposium on Discrete Algorithms (SODA 2014).
He was also recognized for his work addressing another open question central to the field of combinatorial optimization. Multiple authors had already established that various polytopes have exponential extension complexity for NP-hard problems. However, Rothvoss wanted to see if the same was true for polytopes that admit polynomial time algorithms to optimize linear functions. In a paper titled “The matching polytope has exponential extension complexity,” Rothvoss was able to prove that linear programming cannot be used to solve the matching problem in polynomial time, advancing theoreticians’ understanding of the topic. He earned both the 2018 Delbert Ray Fulkerson Prize and the 2023 Gödel Prize for the work — the former honoring exceptional papers in discrete mathematics, the latter for outstanding papers in theoretical computer science
More recently, Rothvoss and his former Ph.D. student Victor Reis (Ph.D. ‘23) utilized geometric tools to develop a faster new algorithm for solving integer programming problems within a fixed set of variables. Integer programming is used for optimizing problems with whole number amounts, but many of the algorithms previously introduced for these types of problems have been quite slow based on the number of steps required. Instead, the researchers resolved a version of the Subspace Flatness Conjecture, and proved a new upper bound on the time required to solve for any integer program. While the breakthrough is theoretical, the research has opened up new questions for theoreticians to pursue, earning Rothvoss a Best Paper Award at the 64th IEEE Symposium on Foundations of Computer Science (FOCS 2023).
In the 1995 movie “Clueless,” lead character Cher Horowitz has a digital closet that allows her to virtually try on outfits. Proving some ideas never go out of style, Allen School professor Ira Kemelmacher-Shlizerman has spent the past two decades working to make that and other futuristic technologies a reality.
“My general area is at the intersection of computer vision and graphics, or generative media. I am excited about all aspects of image and video generation and the applications we can build on top of it. A big passion of mine is to model people and clothing from large photo collections,” said Kemelmacher-Shlizerman, director of the UW Reality Lab and member of the UW Graphics & Imaging Group (GRAIL). “I am currently working to push the boundaries of human and clothing modeling to new levels, going to extreme quality and details.”
The Institute of Electrical and Electronics Engineers (IEEE) recently recognized her in the 2026 class of IEEE Fellows, one of the organization’s highest honors, for her “contributions to face, body, and clothing modeling from large image collections.” The IEEE Fellows represent members with an exceptional record of accomplishments in their field who bring the “the realization of significant value to society at large.”
Kemelmacher-Shlizerman pioneered the field of face modeling from large photo collections in the wild. She helped develop the first photometric stereo method that can reconstruct three-dimensional face models from unstructured photos pulled from the Internet, launching thousands of follow-up papers. As the field progressed, Kemelmacher-Shlizerman and her collaborators introduced the MegaFace Benchmark, the first million-image scale image dataset for evaluating facial recognition algorithms. The work served as the field’s standard benchmark for many years. Kemelmacher-Shlizerman then developed a personalized image search engine that allows users to imagine how they could look with various hairstyles, or in different time periods or ages — anything that can be queried in the search engine. She later commercialized the technology through her startup Dreambit, which was acquired by Meta.
Since then, Kemelmacher-Shlizerman has transitioned from helping users test out new hairstyles to trailblazing virtual try-on technology, allowing for real clothing to be rendered on a human body. She and her team introduced the first use of conditional generative adversarial networks for photorealistic try-on. In addition to virtual try-on technology for photos, Kemelmacher-Shlizerman developed Fashion-VDM, a video diffusion model for generating virtual try-on videos to help users see the garment from multiple angles and understand how it flows and drapes in motion.
“We are working on detailed measurement of humans and clothing to enable fit aware virtual try-on, as in going beyond generative visualization to providing metrically correct measurement based try-on to create better than physical try-on technology,” said Kemelmacher-Shlizerman.
Her work as a principal scientist at Google, where she leads Google Shopping’s Generative Media team, is bringing this technology to the mainstream. Kemelmacher-Shlizerman and her team introduced a generative AI tool for Google Shopping where users can see how an article of clothing looks on a range of models of different sizes, body shapes and skin tones. Last December, her team launched an upgraded version of the virtual try-on tool that generates a full body digital version of a user to help them see how clothes would look like on their own body.
“Ira is arguably the foremost researcher in the field of virtual try-on technology. She has written several seminal papers, pioneering the use of generative AI for this use case, and is the chief technologist behind Google’s launch of this technology in their search and shopping products,” said Allen School professor Steve Seitz, who co-directs the UW Reality Lab. “While this research area is relatively new, it’s already having a major impact on industry.”
In addition to being named an IEEE fellow, Kemelmacher-Shlizerman has been recognized as an Association for Computing Machinery (ACM) Distinguished Member and received a Google Faculty Award, GeekWire Innovation of the Year Award, and the Madrona Prize.
From left to right: EMNLP program chairs Violet Peng and Christos Christodoulopoulos, lead author Hao Xu, EMNLP general chair Dirk Hovy and program chair Carolyn Rose.
Large language models (LLMs) such as ChatGPT are trained using massive text datasets downsampled from the Internet. As these language models become more popular and widespread, it becomes increasingly important to understand the composition of the data source and how it affects the model’s behavior. The first step is to make these texts searchable.
Current exact-match search engines are limited by their high storage requirements, hindering their application on extremely large-scale data. With previous methods, storing the Internet-size text index would cost around $500,000 per month. To make searching on such a large scale more efficient and affordable, a team of University of Washington and Allen Institute for Artificial Intelligence (Ai2) researchers developed infini-gram mini, a scalable system that uses the compressed FM-index data structure to index petabyte-level text corpora.
“We developed infini-gram mini, an efficient search engine designed to handle exact-match search on arbitrarily long queries across Internet-scale corpora with minimal storage overhead,” said Allen School undergraduate student and lead author Hao Xu. “Infini-gram mini hosts the largest body of searchable text in the open-source community.”
While the FM-index has been widely used in bioinformatics, the team was the first to apply it to natural language data at the Internet scale. The infini-gram mini system improves on the best FM-index implementation, achieving an 18 times increase in indexing speed and a 3.2 times reduction in memory usage. The resulting index needs only 44% as much storage as the raw text, which is only 7% of what the original infini-gram required.
“In infini-gram mini, we combined advanced algorithms and data structures and scaled-up engineering to tackle real, pressing challenges in AI. It is a very unique combination,” said co-author and Allen School Ph.D. student Jiacheng Liu. “The most interesting part was that we revitalized a data structure repo that hasn’t been maintained for almost 10 years, armed it with modern parallel computing, and scaled it up to the sky to handle Internet-scale data with low compute needs. We almost built a Google Search without Google-level budget.”
To showcase infini-gram mini’s search capabilities, the researchers used the system to conduct a large-scale benchmark contamination analysis to see if the training data of LLMs inadvertently contains their test data. They found that many widely-used evaluation benchmarks appeared heavily in these corpora, which could lead to an overestimation of the language model’s true capabilities as it enables models to retrieve memorized answers from training data rather than performing task-specific reasoning. Alongside infiini-gram mini, the team also released a benchmark contamination monitoring system, with the goal of encouraging more transparent and reliable evaluation practices in the community, explained Xu.
Additional authors include Allen School professors and Ai2 researchers Hannaneh Hajishirzi and Noah A. Smith, along with Allen School affiliate faculty member Yejin Choi, faculty member at Stanford University.
