New Faculty Highlights

AAAI is continuing its invited speaker program by highlighting AI researchers who have just begun careers as new faculty members or the equivalent in industry. 

New Faculty Highlight talks will be allotted 20 minutes each; the aim for these talks is to broadly survey the candidate’s research to date. Several talks will be released online and publicized each day of the AAAI conference, following which they will be available archivally as part of the conference program. Invited speakers will be further invited to contribute an article to a corresponding series in AI Magazine. 

Program Overview

Towards Agents that Exhibit Human-Like Autonomy in Complex Environments

Rohan Chandra

Deploying intelligent, autonomous agents e.g. autonomous vehicles and robots, in the real world has been a longstanding goal in robotics and artificial intelligence (AI). We have already begun to witness the emergence of vacuum robots in our homes, service robots in warehouses, and even self-driving cars on our way to work. These environments are often dense, constrained, and unstructured, with heterogeneous agents, each with their own unique behaviors and objectives. While agents today are designed to navigate these environments safely, their overly conservative nature often leads to slow and jerky motion (frequent stopping and freezing), lack of social compliance (not giving way to other people, blocking doorways and intersection), and poor adaptability across diverse complex environments (failure due to sudden accidents e.g. liquid spills). In other words, these robots often fail to capture the essence of human-like autonomy, which involves the ability to take calculated risks, even in complex environments. In this talk, I will describe my vision for a paradigm shift in the way intelligent physical agents navigate highly dense, heterogeneous, constrained, and unstructured environments using human-like autonomy. 

From Few-Shot Learning to Data-Efficient Intelligence

Yaqing Wang

Modern artificial intelligence performs impressively in data-rich settings but still struggles to learn and adapt from only a few examples—a capability central to human intelligence. My research seeks to understand and enable data-efficient generalization, unifying principles across few-shot learning, meta-learning, in-context learning in large language models (LLMs), and adaptive agent behavior. First, I revisit few-shot learning from a foundational perspective, showing why conventional supervised learning breaks down under sparse data and how prior knowledge enables reliable adaptation. I then discuss how these principles extend to real-world scenarios such as scientific discovery and cold-start recommendation, where data are scarce, costly, or dynamically evolving. Finally, I explore how LLMs perform in-context learning and how their adaptive behaviors connect to meta-learning mechanisms. Building on these insights, I develop data-efficient, preference-adaptive agents that quickly align to user needs with minimal interaction.This talk presents a cohesive view of data-efficient intelligence and outlines future directions toward more reliable, human-like learning systems.

Toward Trustworthy AI for Decision Making in Population Health

Alexander Rodríguez

AI and public health are becoming increasingly intertwined, driven by the growing availability of multimodal data and rapid advances in deep learning. In this talk, I will present our efforts to harness these trends within a data-centric pipeline for public health. I will first discuss robust deep learning architectures for real-time outbreak response, highlighting how our frameworks capture uncertainty and dynamics across shifting distributions, multimodal signals, hierarchical structures, and relational dependencies. I will then introduce hybrid approaches that integrate machine learning with mechanistic epidemiological models, including physics-informed neural networks, expert-guided generative models for causal inference, and differentiable agent-based models. Together, these advances illustrate how combining data-driven AI with domain knowledge can enable more reliable, adaptive, and actionable public health solutions for decision making.

Scaling Human-Centric Trustworthy Foundation Model via Advanced Reasoning and Agentic Frameworks

Yi R. (May) Fung

As foundation models grow in size and scope, crucial challenges remain in scaling their trustworthiness and adaptability to meet the diverse needs of individual users, as well as mitigating their risk of generating unhelpful, non-factual, or harmful content. To address this, we propose to reframe model reasoning through a unified paradigm of active knowledge grounding that coordinates different tools and modalities. First, to scale reasoning depth and creativity, we introduce the novel paradigm of Thinking with Images to encourage models to externalize intermediate structure and perform interleaved cross-modal advanced reasoning beyond text-centric cues. To further scale honesty and bridge knowledge gaps reliably, we develop one of the first vision-language deep research agents, WebWatcher, that actively gathers and verifies information from the web with enhanced fragmented reasoning capability. Ultimately, to scale effective and efficient human-AI collaboration, we propose AdaCtrl as a novel training mechanism for dynamically aligning model behavior with individual user preferences and difficulty awareness to adaptively allocate computational resources. Together, these three pillars of integrating advanced multimodal reasoning, autonomous discovery, and adaptive alignment form a foundational framework for advancing the frontier of next generation human-centric trustworthy AI systems.

Advancing Trust in Multimodal AI: From Explainability Benchmarks to Clinical Safety

Chirag Agarwal

Machine learning models have become ubiquitous in the last decade, and with their increasing use in critical applications (e.g., healthcare, financial systems, and crime forecasting), it is vital to ensure that ML developers and practitioners understand and trust their decisions. This problem has become paramount in the era of frontier models, which are developed by training billion-parameter models on broad, uncurated datasets and extensive computing. In this talk, we will first explore the (un)reliability of existing multimodal explainability techniques in large language and multimodal models and understand the robustness and safety implications of Mechanistic Interpretability tools. Next, we will delve into two complementary threads: i) domain-specific safety and related trustworthy evaluation that surfaces risks missed by generic red-teaming, focusing on multilingual and distribution-shifted settings; and ii) methods that explicitly train and assess reasoning in medical LLMs.

Toward Controllable and Trustworthy LLM Reasoning: From Failure Mapping to Cognition-inspired Control and Real-world Impact

Ben Zhou

Large language models (LLMs) are increasingly deployed in decision-critical settings, yet still exhibit brittle, opaque reasoning, especially under abstraction, spurious cues, and long-horizon planning. In this talk, I will present my research on making LLM reasoning controllable and trustworthy. First, I will show how to map when and why LLM reasoning fails, revealing conceptual blind spots, deceptive semantic shortcuts, and brittle consistency that undermine trust. Second, I will introduce cognition-inspired methods that steer model reasoning via problem and factual decomposition, consistency-driven learning, and controlled self-reflection, as well as pre-training schemes that attach explicit “thoughts of words” and token-level reinforcement learning. Finally, I will demonstrate how trustworthy reasoning can be translated into real-world impact, closing the loop between failure analysis, controlled reasoning, and safe AI applications.

Towards faithful, interpretable and private foundation models 

Grigorios Chrysos

As large language models (LLMs) scale, ensuring interpretability and privacy becomes critical. This talk addresses these interconnected challenges with novel approaches to model specialization and safety. First, we tackle the dense, distributed nature of LLM representations by casting Mixture-of-Experts as a tensor decomposition, enabling specialized experts in a factorized space. Second, we argue that current neuron-level sparsity methods create a severe accuracy-sparsity trade-off, and we propose a paradigm shift to layer-level sparsity with the Mixture of Decoders. We explain how MxD uses tensor factorization to expand dense layers into thousands of specialized, full-rank sublayers. Finally, we address privacy in open-weight models by proposing a scalable alternative to differential privacy that induces maximal uncertainty on protected instances, introducing a certifiable algorithm and proving tight bounds that characterize the resulting privacy-utility tradeoff.

KOALA: Knowledge of Optimization and Learning Algorithms for Healthcare

Kai Wang

The Knowledge of Optimization And Learning Algorithms (KOALA) group studies how to integrate optimization, machine learning, and generative modeling to enable data-driven decision-making under uncertainty. We study decision-focused learning, embedding optimization as a differentiable layer to train models end-to-end for decision quality. We design scalable reinforcement learning algorithms for population and personalized healthcare, and develop efficient bilevel optimization methods for nested and multi-agent decision-making. These directions form a unified framework linking optimization and learning for impactful AI in healthcare. Through collaborations with hospitals and NGOs, our group designs and deploys algorithms for pediatric, diabetes, maternal, and mental health applications. Looking ahead, we aim to unite these foundations with generative AI to build theoretically grounded and socially responsible algorithms that advance trustworthy, real-world AI for health and beyond.

Diffusion-Based Data Augmentation for Bimanual Robot Manipulation

Daniel Seita

Learning robust bimanual manipulation policies requires demonstration data with broad coverage over robot poses, contacts, and scene contexts. However, collecting diverse real-world demonstrations is costly and time-consuming, creating a significant bottleneck for scaling imitation learning systems. This talk presents two complementary approaches that leverage diffusion models to address this challenge. First, D-CODA synthesizes novel viewpoint-consistent wrist camera images for eye-in-hand bimanual setups, and uses constrained optimization to generate action labels that satisfy coordination constraints during contact-rich manipulation. Second, ROPA extends this framework to third-person setups with RGB-D observations, introducing skeleton-based pose conditioning to generate diverse robot configurations while maintaining geometric consistency across both RGB and depth modalities. Experiments in simulation and real-world bimanual tasks demonstrate that these methods substantially outperform baselines, enabling more sample-efficient imitation learning for coordinated manipulation without requiring additional human demonstrations or simulator access.

Deep Model Reuse: Toward Flexible and Efficient Generative AI

Xingyi Yang

Humans easily apply learned skills to different situations, a flexibility that AI systems still struggle to achieve. Despite the explosion of large, specialized models, most are trained and deployed in isolation, resulting in redundant computation, brittle generalization, and limited adaptability. Deep Model Reuse offers a paradigm shift: instead of training from scratch, we treat the growing ecosystem of pre-trained models as a dynamic knowledge library. By strategically extracting, aligning, and re-composing their latent capabilities, we can rapidly construct new, versatile AI systems with minimal additional cost.
In this talk, I will present key techniques that enable deep model reuse, and demonstrate their transformative impact in generative and multimodal AI. From controllable image synthesis to 3D content generation, our approach bridges isolated model islands into a coherent, adaptable intelligence. Ultimately, deep model reuse moves us closer to AI that learns like humans: efficiently, flexibly, and cumulatively.

Teach AI What It Doesn’t Know

Sean Du

The remarkable capabilities of machine learning (ML) models, especially foundation models like GPT, have transformed numerous domains. However, these systems often falter in real-world settings, where they encounter unknown or out-of-distribution (OOD) inputs, and generate overconfident predictions or unreliable outputs. Ensuring their reliability is not only a technical challenge but also a fundamental requirement for their safe deployment.
In this talk, I will discuss my research on teaching ML models what they don’t know by developing foundational frameworks for reliable decision-making in the open world. This involves three core aspects: (1) designing novel algorithms for unknown-aware learning through adaptive outlier synthesis, enabling models to handle unfamiliar inputs without explicit knowledge of unknowns; (2) addressing reliability blind spots in foundation models, such as hallucinations and malicious prompts, through innovative mitigation strategies; and (3) rethinking the AI alignment process from algorithm-centric to data-centric research that enables learning better reward models from limited human preferences. 
Through fundamental algorithmic development, theoretical insights, and practical applications, my research contributes to the responsible deployment of AI technologies. The talk will conclude with a forward-looking perspective on interdisciplinary collaborations and the roadmap for achieving robust, reliable AI systems that adapt to an ever-changing world.

Trustworthy Planning with Large Language Models Generating Formal Languages

Li “Harry” Zhang

Despite the rapid advancement of AI, most systems in high-stakes applications remain primarily limited to rule-based interactions and cannot reliably plan or execute complex user tasks. Despite recent efforts in using large language models (LLMs) to plan as agents, their hallucinations and lack of verifiability undermine executability and trust, preventing real-world deployment. This proposal advances an alternative paradigm: LLM-as-formalizer. Instead of relying on LLMs to generate plans directly, we use them as a code generator to translate a user’s environment and goal into formal languages (such as PDDL) that can be deterministically solved by off-the-shelf solvers. This neuro-symbolic approach combines the flexibility of LLMs with the reliability of symbolic systems, offering a pathway toward trustworthy, generalizable planning. Based on our published prior work that has demonstrated the feasibility and challenges of the LLM-as-formalizer approach through comprehensive evaluation across multiple domains and models, I propose to significantly advance this paradigm by exploring ways to improve LLMs’ formalizing ability in various domains.

Learning from Imperfect Data: Continual Learning, Few-shot Learning, and Generative Data Augmentation

Yaoyao Liu

In recent years, artificial intelligence (AI) has achieved great success in many fields. Although impressive advances have been made, AI algorithms still suffer from an important limitation: they rely on static and large-scale datasets. In contrast, human beings naturally possess the ability to learn novel knowledge from real-world imperfect data, such as a small number of samples or a non-static continual data stream. Attaining such an ability is particularly appealing and will push the AI models one step further toward human-level Intelligence. In this talk, I will present my work on addressing these challenges in the context of continual learning, few-shot learning, and generative models for data augmentation. First, I will discuss how to get better exemplars for continual learning based on optimization. I parameterize exemplars and optimize them in an end-to-end manner to obtain high-quality memory-efficient exemplars. Next, I will introduce our work on generative models that leverage diffusion techniques to create diverse, 3D-annotated images, enabling large-scale 3D model training without human annotations. Finally, I will present my work on how to apply continual and few-shot learning techniques to more challenging and realistic scenarios, e.g., object detection and medical imaging.

Towards Inclusive AI: Advancing Multilingual Large Language Models

Dr. Wenxuan Zhang

Large language models (LLMs) have advanced rapidly, yet their development remains disproportionately focused on a few high-resource languages, leaving fundamental scientific and societal questions about multilingual capability, safety, and equity unresolved. This talk examines multilingual LLMs as a lens for understanding these challenges. I will first discuss observations from large-scale evaluations with real-world natural data, which reveal substantial performance gaps and highlight the need to treat multilingualism as a multidimensional construct. I then turn to safety, presenting work that uncovers multilingual jailbreak vulnerabilities and introduces frameworks for achieving more consistent cross-lingual alignment. Building on analyses of language-specific internal mechanisms, I will outline new strategies for enhancing multilingual systems and describe open-source efforts such as the SeaLLMs and Babel projects that aim to broaden linguistic and cultural inclusivity. Finally, I will discuss emerging directions beyond language, including recent findings on abstract thought in LLMs, which point toward the development of models that are not only multilingual but genuinely multicultural and contextually grounded.

Reinforcement Learning Without Explicit Rewards: Theory and Practice

Weitong Zhang

The deployment of reinforcement learning in modern systems often relies on implicit feedback rather than a precisely specified reward. This creates two central challenges. First, exploration tied to task rewards can be inefficient when rewards vary or are misspecified across tasks. Second, preference or proxy rewards can invite reward hacking. My research addresses these issues by advancing reinforcement learning without explicit reward design, with theoretical understanding and practical algorithms. In this talk, I will first discuss unsupervised and reward free exploration. The key idea is to acquire broad coverage of states and skills with curiosity, then finetune the agent with misspecification toleration. Concretely, we introduce a theoretical framework with intrinsic rewards. We then instantiate these ideas with practical algorithms for uncertainty aware exploration and representation learning that show strong empirical performance. Next, I will present methods that leverage the relative reward signals in flow matching. I will discuss exact energy guidance for flow matching where the model can guarantee the exact posterior distribution of the generation process. I will also showcase interdisciplinary applications enabled by these foundations including the energy guided ligand generation. Finally, I will highlight our work that pioneers machine learning for electrochemical analysis. The talk will close with open challenges and future directions, including theory for learning from implicit signals under distribution shift and reliable offline to online adaptation with safety and constraint guarantees.

Breaking the Resource Monopoly: LLM Post-Training and Serving with Modest Data and Compute

Jiaxin Huang

State-of-the-art large language models are increasingly powerful though many of them are trained from vast proprietary data and intensive computes, raising barriers for academic labs and smaller institutions for exploration and improvement. In this talk, I will present a unified research agenda for breaking the resource monopoly in both post-training and serving. On the training side, I will describe label-free and even zero-data post-training pipelines that let models curate their own reasoning supervision and evaluate it with controllable, fine-grained benchmarks. On the serving side, I will show how cost-aware inference can enable adaptive test-time scaling to be more efficient, by adjusting computing resources between queries of various difficulty levels. In addition, I will also introduce how to make language model generation more reliable by calibrating post-trained models to be less overconfident. Together, these components form a practical blueprint for building capable, trustworthy, and efficient LLM systems using modest data and compute resources.

Towards Continually-Evolving AI: Selective and Expandable Multimodal Memory

Jaehong Yoon

This talk presents a new paradigm for continually-evolving AI, where agents selectively expand their knowledge and build multimodal memories that support long-horizon reasoning. I will introduce how selective learning and expandable memory enable AI to refine skills, acquire new ones, and interpret complex visual experiences over time. First, I will discuss continual multimodal instruction tuning through the lens of training data selection. Visual instruction datasets often arrive asynchronously and contain substantial redundancy, making continual improvement inefficient. The proposed approach implicitly groups samples by their underlying skills and identifies the examples that most effectively strengthen the model’s current knowledge while still supporting steady refinement and the acquisition of new abilities. Next, I will introduce a multimodal memory framework for long-form video understanding. While video language models perform well on short clips, they struggle with hours-long content where important visual details disappear through abstraction and events unfold across diverse timescales. In contrast, we build complementary episodic, semantic, and visual memories and retrieve them adaptively, ensuring that both high-level concepts and fine-grained visual evidence remain accessible during reasoning.

Cracking the Code of Health Communication: Explanation, Evaluation, and Personalization

Yue Guo

Although scientific papers contain the information needed to support health-related decision making, this knowledge is often not accessible to the public. Medical jargon, dense writing styles, and missing background explanations make biomedical information opaque to non-experts. This creates an urgent need for scientific knowledge to be communicated in clear, actionable language. In this talk, I present my research on plain language summary (PLS) generation–automatically translating science texts into lay-friendly summaries. I will discuss four key challenges and my contributions toward addressing them: building large-scale training data, generating effective explanations, developing reliable evaluation methods, and personalizing summaries to readers’ backgrounds.

Safe Reinforcement Learning for Trustworthy AI: Theory, Algorithms, and Applications

Honghao Wei

Safe reinforcement learning (RL) has emerged as a key paradigm for deploying AI in high-stakes domains such as autonomous driving, robotics, healthcare, and recommender systems. By embedding constraints into the learning process, safe RL enables agents to optimize performance while satisfying critical requirements, including collision avoidance, resource limits, and system reliability. Such guarantees are indispensable for real-world AI, where failures can cause physical harm, economic loss, or loss of trust. At the same time, demand for trustworthy AI continues to grow as machine learning is increasingly deployed in human-centered applications. This makes it essential to design RL algorithms that are not only efficient but also reliable, robust, and aligned with societal needs.

Reasoning LLMs for Science and the Science of Their Reasoning

Hao Peng

In this talk, I will first focus on our recent works using reasoning-capable LLMs to accelerate scientific discovery, highlighting new benchmarks and evaluations that investigate their effectiveness in realistic research workflows. I will then turn to the science of reasoning LLMs themselves, drawing on our recent studies that investigate the mechanisms underlying their reasoning abilities, how these can be elicited or enhanced, and what this reveals about their potential and limitations. Together, these perspectives offer a unified view of LLMs both as engines for advancing science and as subjects of scientific inquiry.

All-Purpose Mean Estimation over ℝ

Jasper Lee

Given society’s increasing reliance on data, its collection and processing into useful information is a technical problem of growing focus, and perhaps paradoxically, a critical bottleneck in many data science and machine learning applications. Yet, even for the most basic statistical problems such as mean estimation, there is a theory-practice divide. Conventional methods like the sample mean, while supported by theoretical results under strong assumptions, are often brittle in the presence of extreme data. Practitioners thus often use ad-hoc and unprincipled “outlier removal” heuristics, but which can lead to wrong conclusions (e.g. Milikan’s underestimation of the electron charge (Holton 1978)).
In this talk, I will describe my work that essentially resolves the fundamental 1-d mean estimation problem. I will show the construction of a statistically-optimal and computationally-efficient 1-dimensional mean estimator, whose estimation error is optimal even in the leading multiplicative constant, under bare minimum distributional assumptions (FOCS 2021). Furthermore, I will discuss its various robustness properties (ICML 2025 Oral), in particular highlighting robustness to adversarial sample corruption.

Cross-Modal Knowledge Transfer in Time Series AI via Large Vision Models

Jingchao Ni

Time series analysis has progressed from traditional autoregressive models to deep learning, Transformers, and foundation models (FMs), including large language models (LLMs) and large vision models (LVMs). These advances have expanded model design possibilities and, notably, enabled time series problem-solving across multiple modalities, greatly improving downstream applications in domains such as climate, energy, and healthcare. This talk will provide an overview of recent developments in large FMs for time series, highlighting frameworks for transferring knowledge from other modalities to time series, and identifying the advantages of LVMs over LLMs in cross-modal knowledge transfer. I will then delve into our recent research on LVMs for time series, discussing (1) mainstream techniques for imaging time series; (2) key strengths and limitations of LVMs in time series modeling; and (3) multimodal frameworks that integrate LVMs for time series encoding. This talk will conclude with applications and future directions. The aim of the talk is to review state-of-the-art AI techniques for time series, highlight unique challenges, and share our recent findings in this promising area.

Augmenting Human Creativity with Machine Learning

Hao-Wen (Herman) Dong

In this talk, I will survey my work in three main research directions: 1) generative models for music creation, 2) AI-assisted music creation tools, and 3) multimodal generative models for content creation. In particular, I will discuss our recent work on AI-assisted video editing that explores novel machine learning models that can cut, select, and rearrange a long video into a short video. In the first TeaserGen project, we proposed a narration-centered teaser generation system that can effectively compress >30-min documentaries into <3-min teasers leveraging pretrained LLMs and language-vision models. In the second REGen project, we proposed a retrieval-embedded generation framework that allows an LLM to quote multimodal resources while maintaining a coherent narrative. I will conclude by discussing our future work towards next-generation video editing interfaces using multimodal LLMs and retrieval embedded generation. I will also discuss our future work towards playful human-AI music co-creation systems where the user can control a music generation system through hand gestures and body movements.

Towards Human-centered Proactive Conversational Agents

Deng Yang

Conversational AI agents are envisioned to provide social support or functional service to human users via natural language interactions. The popularity of conversational AI has grown unprecedentedly with the advent of ChatGPT, which showcases exceptional proficiency in the capabilities of context understanding and response generation with large language models (LLMs). However, typical conversational systems are built to follow instructions, which means that the conversation is led by the user, and the system simply follows the user’s instructions or intents. My research endows the conversational AI with the capabilities of creating or controlling the conversation to achieve the conversational goals by taking initiative and anticipating impacts on themselves or human users (Intelligence), namely Proactive Conversational AI. I will also highlight the importance of moving towards building human-centered proactive conversational AI that emphasize human needs and expectations (Adaptivity), and that considers ethical and social implications of these agents (Civility), rather than solely focusing on technological capabilities.

Efficient Model Specialization via Training-time and Test-time Adaptation

Huanrui Yang

The scaling up of single model size is stalling due to the inadequate high-quality data to pretrain on and the diminishing return in performance gain. Inspired by the trend of specialization and heterogeneity explored in the semiconductor’s Moore’s Law, we discuss effective methods to specialize a pretrained model to heterogeneous variants for both improved efficiency and performance. In this talk, we discuss efficient model specialization algorithm to adapt the pretrained model towards downstream tasks while improving its efficiency, efficiently generalizing to multiple tasks via dynamic architectures, and improving inference-time efficiency utilizing the diversity within model block functionalities. These research directions serve as the foundation towards co-designing models, tasks, systems, and hardware for a reconfigurable efficient intelligence future.

Bridging AI with Clinical Decisions from a Data Centric Perspective

Jiaming Cui

Machine learning and clinical decision-making are intertwined. ML methods depend on high-quality clinical data and domain knowledge to produce accurate and robust predictions, while clinical decisions increasingly rely on ML outputs to guide optimal interventions. Dr. Cui’s research brings a data-centric perspective to bridge AI with clinical decision-making. His work addresses multiple challenges arising from effectively utilizing multimodal clinical datasets and issues stemming from the complexity of disease spread dynamics in healthcare facilities. This talk will cover methods developed to address these challenges with better designed models to optimize disease surveillance and control policies and new techniques for end-to-end learning with mechanistic models. The talk will conclude by discussing emerging challenges and opportunities at the intersection of machine learning, scientific modeling, and clinical decision-making for computer scientists, epidemiologists, and computational biologists.

Towards Aligned and Efficient Large Language Models

Yu Meng

Large language models (LLMs) have rapidly transformed the landscape of AI, demonstrating remarkable capabilities across reasoning, communication, and problem-solving. Yet, realizing their full potential requires addressing two critical challenges. First, their behavior must be steered and refined after training to ensure reliability, safety, and alignment with human values and intentions. Second, their large scale comes with substantial costs in training and deployment, necessitating research into more efficient methods.
My research centers on advancing both of these fronts—making LLMs both aligned and efficient. On one side, I investigate post-training techniques that allow models to better reflect human preferences, demonstrate strong reasoning capabilities, and mitigate hallucination. On the other side, I study methods for improving data efficiency in training and inference efficiency in deployment. Together, these thrusts highlight a broader vision of enabling LLMs that are not only powerful, but also trustworthy and accessible at scale.

Robust Machine Learning for Biomedical Data: Efficiency, Reliability, and Generalizability

Chenyu You

In the rapidly growing area of machine learning, there is profound promise in crafting intelligent, data-driven methods for diverse real-world applications. Yet, in safety-critical domains like healthcare, some fundamental challenges remain: (1) The insufficiency of raw biomedical data emphasizes the need for data-efficient and robust learning approaches. (2) The imperative of safety and stability necessitates a cohesive framework that unifies learning with theoretical guarantees. (3) The inherent heterogeneity and distribution shifts in real-world clinical data call for robust and generalizable learning methods. To address these challenges, there are several major directions I have explored: (i) (Robust) Machine Learning for Imperfect Medical Data: The development of machine learning models, particularly in the context of label scarcity, increasingly necessitates the collection of substantial annotated medical data. Moreover, medical data often display a long-tailed class distribution, which consequently results in notable imbalance issues. To this end, there are several growing interests in training machine learning models jointly across imbalanced class distributions and limited annotations. I have developed novel, efficient, statistically consistent algorithms to improve empirical performance for biomedical image analysis. (ii) Learning with Theoretical Guarantees: As machine learning methods have become ubiquitous in clinical decision-making, their reliability and interpretability have become important. This is particularly crucial in the field of biomedical image analysis, where decision outcomes can have profound implications. I have developed novel machine learning algorithms that enable provably accurate anatomical modeling with theoretical guarantees. (iii) Generalize across Diverse Biomedical Data: The development of medical foundation models often requires massive and diverse biomedical data. To this end, I have developed various foundation models for biomedical imaging data and explored novel applications of these models. I have also developed novel medical AI Agents that lead to the scalable and accurate predictive modeling, particularly for distribution shift problems.

Graph-based Label-Efficient Learning: When Graph-Structured Data Meets Limited Labels

Zixing Song

The success of deep learning is highly dependent on large-scale labeled data. This dependency creates a challenging bottleneck in high-stakes scientific domains, such as molecular discovery, where data annotation is prohibitively expensive. Consequently, developing label-efficient learning methods to maximize model performance under limited annotation budgets has recently become more and more critical. However, mainstream label-efficient strategies still suffer from a Euclidean bias. Standard active learning or semi-supervised methods primarily operate on grid-structured or sequential data. They fail to effectively capture the non-Euclidean, or relational dependencies inherent in graph-structured data. In this talk, I present a systematic framework to bridge this gap, outlining a progressive path toward label-efficient graph machine learning. First, we address the adaptation of general label-efficient paradigms to explicit graph data. We demonstrate how to strategically select labeled nodes to maximize performance gain, how to effectively utilize unlabeled nodes to capture topological priors, and how to synergistically combine both labeled and unlabeled nodes to enforce prediction consistency. Second, we explore the utilization of implicit graph data to enhance general label-efficient learning. We pioneer methods to construct and leverage latent graph structures within unstructured data, propagating label information to significantly boost prediction performance.

Toward Causal Foundation World Models: From Representation to Decision-Making  

Mengyue Yang

World models are becoming a central interface between perception, prediction, and control, yet most current systems are fundamentally correlational and assume fixed dynamics, limiting their robustness in the non-stationary, open-ended domains we care about (robots, science, society). In this survey, I will outline a research agenda for causal foundation world models—scalable, general-purpose models that embed causal structure to support explanation, counterfactual reasoning, and decision-making. I will briefly connect recent advances in causal representation learning, LLM-based reasoning and planning, and multi-agent decision-making with emerging ideas on curiosity-driven intervention and dynamic causal discovery, and argue that bringing these strands together is key to building trustworthy world models that can generalize, adapt, and provide meaningful guarantees in real-world applications.

From Representation to Reasoning: Toward General-Purpose Visual Intelligence

Chen Wei

Visual perception is our primary interface with the world, yet today’s AI remains largely confined to static recognition. My research asks a more fundamental question: What is the structure of visual experience, and how can machines internalize it as a basis for understanding and action? By uncovering how visual signals map to concepts, language, and imagination, we build representational and generative foundations that allow models not only to perceive what is present, but to infer what is missing, and to hypothesize what could be. From this foundation, I explore how vision can become a vehicle for reasoning. PyVision enables models to construct and deploy their own tools for problem-solving, while ViGaL uses gameplay to incentivize cognitive skills that transfer beyond any single task. These efforts push AI from passive understanding toward agents that can explore, anticipate consequences, and engage with the world on their own terms. The goal is an AI that doesn’t merely see, but thinks through vision.