Abstracts 

Generative AI has entered scientific research at unprecedented speed, and the productivity evidence is considerable: significant increases in publication output, higher citation rates for AI-assisted papers, and measurable gains concentrated among less-experienced researchers and resource-constrained institutions. Yet these individual benefits do not translate automatically into system-level improvements. Roughly 95 percent of AI use in scientific publishing goes undisclosed, undermining merit-based evaluation. Democratization gains remain unevenly distributed across geographies, genders, and disciplines. Experimental evidence suggests that passive AI use compromises learning, while critical engagement enhances it, raising concerns about doctoral training in the absence of institutional guidance. This keynote examines the tension between AI's documented benefits and the governance gap that accompanies its rapid adoption. Drawing on recent empirical evidence, I argue that self-regulation is not converging toward effective standards, and that the research community needs a governance architecture capable of preserving AI's gains while addressing the selection, training, and distributional risks that individual researchers cannot manage alone.

How do biological and artificial systems retrieve a stored pattern from a noisy cue? The Hopfield network frames this as a problem of energy minimization, where memories are stored as energy minima. While classical models have limitations, modern generalizations achieve exponential memory capacity through a retrieval rule mathematically identical to the "attention mechanism" in Transformer architectures. This talk will outline mathematical principles behind these models and explore how the energy-landscape perspective provides a unified framework for understanding neural circuits, large language models, and complex biological systems.

Symmetry is the bedrock of theoretical physics, upon which our understanding of nature has been built — from classical mechanics to the Standard Model of particle physics. Traditionally, however, symmetry has been synonymous with symmetry groups, which are global and often too rigid to adequately capture the complexity of biological systems. Drawing inspiration from Grothendieck's notion of fibrations in category theory, a natural generalization of symmetry groups can be constructed through a hierarchy of graph symmetries, ranging from fibrations and opfibrations to coverings. This hierarchy extends the strict symmetry groups of physics to local, more flexible forms, which prove fundamental in neuroscience: the multiscale organization of neurons gives rise to a broad spectrum of hierarchical functions, posing a longstanding challenge in systems neuroscience — namely, understanding how local structure relates to function. These local symmetries also shed light on how artificial neural networks operate during both inference and learning, emerging as key organizing principles of their behavior. The aim of this talk is to introduce the framework of graph symmetries and illustrate their applications to neuroscience and deep learning.

The technologies most profoundly influencing contemporary society are based on the Web, on the social networks, and on Artificial Intelligence (AI). While these systems have considerably expanded human capacities for communication and knowledge exchange, they have simultaneously undermined privacy, concentrated control within a few dominant entities, and imposed substantial energy demands. The coexistence of humans and AI remains uncertain and largely unregulated, introducing risks that extend beyond strictly technological domains. In this talk I advocate for a fundamental re-examination of these paradigms, grounded in the principles of privacy preservation, energy efficiency, and distributed system design. I propose a new conceptual model of the Web, one that transitions from a network of “document-like” resources to a network of agents—both human and artificial. In this model, intelligence is relocated from centralized cloud infrastructures to edge devices endowed with increasing local computational capabilities, while retaining the integrative information-sharing dynamics characteristic of contemporary social platforms.

A major assumption is that of re-organizing the Web around communities that we call ``Worlds.'' They are autonomous ecosystems of intelligent agents that function as societies organized around shared goals, topics, or intents. Within these Worlds, agents learn, reason, and plan not only to advance individual objectives but also in response to collective constraints, operating through a peer-to-peer communication protocol. This vision is instantiated in UNaIVERSE, a platform designed to enable decentralized AI and human–AI coexistence, with privacy as a foundational design principle (\url{https://unaiverse.io/}). We analyze relationships with related approaches and outline a series of use cases that illustrate how this paradigm may transform problem formulation and solution methodologies in decentralized intelligent systems.

Before the advent of LLMs, a common pedagogical strategy was to assign students to implement familiar algorithms or modeling procedures from scratch. The underlying functions already existed in standard libraries, but students were asked to build them independently.

These "code-up" exercises reinforced several learning outcomes: they required students to engage with methods at a level of precision that passive use does not demand, they required translation between mathematical formulation and working code, and they encouraged abstraction, since a correct implementation must generalize beyond any single dataset. Crucially, the exercise also developed analytical judgment, particularly as students diagnosed failures and encountered edge cases.  Student understanding came through the simultaneous demands of mathematical and computational thinking.

Widespread LLM adoption has disrupted this pedagogy. When students can generate a working implementation in seconds, the exercise no longer reliably produces the intended struggle. At UVA's School of Data Science, we are responding by asking two questions: which of the underlying competencies remain essential, and how can tasks be restructured to preserve the learning outcomes that still matter? This presentation describes our current thinking on both questions.

The built environment is one of the most consequential yet undertheorized domains for data science and artificial intelligence: buildings and infrastructure account for approximately 39% of global carbon emissions, even as the UN Environment Programme projects roughly 2.5 trillion square feet of new construction by 2060, an amount equal to the entire global building stock in 2023. At the same time, material supply chains and conventional construction systems are under strain from finite resources, shifts in climate, and growing scrutiny of material life cycles. These conditions demand new forms of computation that do not merely optimize existing systems, but help reorient how materials are valued, sourced, processed, assembled, and decommissioned. This talk will explore how data-driven and computational systems can support and enable sustainable construction.

Following AI’s recent developments, both from a technical and from a societal perspective, one often feels like a passenger in a runaway car. As the speed increases, it becomes harder to discern where we are going. We don’t know if the seatbelts will hold or which airbags, if any, will deploy in a crash. Yet the ride is exhilarating and promising for many – and inescapable also for those who are unexhilarated, or simply unaware. As a statistician, I spent the last decades making the most of ever more abundant data and compute. I thoroughly loved developing methods and algorithms, and advocating for data- and computation-driven discovery in science. At this juncture, staying away from doomsday scenarios (e.g., rogue agents, mass surveillance, catastrophic AI war deployment, cybersecurity or bioterrorism risks) and without any pretense to offer an exhaustive critique, I will try to highlight a handful of notions (e.g., friction, enclosement, entry gap, winner-takes-all) that exemplify some key aspects of the current debate, as a means to catalyze discussion during the workshop and subsequent lab.

Artificial intelligence is increasingly framed as a general-purpose technology with the potential to reshape scientific discovery across domains. Yet it remains unclear whether AI is functioning as a broadly diffusing methodological infrastructure or whether its development and application are becoming concentrated within a limited set of countries, institutions, and disciplinary communities. In this presentation, we take a science of science approach to examine the global structure and diffusion of AI knowledge networks. Leveraging large-scale publication, citation, and collaboration data, we trace how AI methods and techniques propagate across scientific domains and national research systems over time.  We characterize the extent to which AI knowledge production is concentrated versus distributed, identify which countries and disciplines serve as major exporters and importers of AI methods, and examine how patterns of collaboration shape the adoption of AI across fields. We further investigate whether AI integration contributes to greater global scientific connectivity or reinforces existing inequalities in access, visibility, and influence. Together, these analyses provide a systems-level view of how AI is reorganizing the structure of modern science and offer insight into whether the emerging AI research ecosystem is defined more by diffusion or isolation.

We study whether personalized learning through large language models improves financial literacy more effectively than alternative digital learning tools. In a preregistered online experiment with 500 UK adults aged 18–40, we compare a general-purpose ChatGPT-style chatbot, a dedicated AI financial tutor, web search, and static text materials, against a Baseline group. All learning interventions significantly increase financial literacy despite high initial knowledge. Differences across tools are modest, with the largest gains observed for the tutor-oriented chatbot. Behavioral effects are more mixed: while interventions reduce allocations to saving instruments when savings are broadly defined, effects on cash holdings and risk diversification are small and mostly insignificant. AI-based tools are evaluated more positively than web search and text, and the AI tutor emerges as the most effective learning environment when learning gains, user experience, and engagement are jointly considered. The findings point to AI tutoring as a promising and scalable approach to financial education, while highlighting the persistent difficulty of shifting financial behavior.

Modern machine learning on graphs is largely built on the assumption that neighboring nodes should be similar, an inductive bias that breaks down in the presence of heterogeneity, multi-modality, and relational ambiguity. Sheaf Neural Networks offer a principled alternative by replacing this notion of similarity with structured consistency: information is no longer directly compared across nodes, but aligned through learnable transformations defined on edges and, more generally, on higher-order relations.In this talk, I introduce the sheaf-theoretic perspective as a natural and flexible framework for learning over complex relational data, spanning both graphs and hypergraphs. I will motivate the approach through concrete failure modes of standard Graph Neural Networks, and show how sheaves provide a unified solution to modeling directionality, heterophily, and context-dependent interactions, while naturally extending to higher-order structures beyond pairwise connections. I will then present a sequence of theoretical results that characterize the expressivity, stability, and spectral properties of Sheaf Neural Networks, highlighting their role as a strict generalization of classical architectures.Finally, I will illustrate how these ideas translate into practice through applications in recommender systems, where modeling structured inconsistency leads to significant empirical improvements. I conclude by discussing open challenges and future directions, including scalable learning of sheaf structures on hypergraphs and connections to broader geometric deep learning paradigms.

Dreams are universal yet highly individual experiences. While memory and personal concerns are known to shape dream content, it remains unclear how these influences evolve over time and how stable individual traits contribute to dreaming. Here, we systematically quantified the semantic structure of dreams using large language model-assisted analyses applied to a multimodal dataset comprising reports of both dreams and waking experiences from 207 adults collected between 2020 and 2024, alongside demographic, cognitive, psychometric, and sleep-related measures. Compared with waking reports, dreams shifted from self-referential, thought-oriented narratives toward perceptual experiences marked by vivid visuo-spatial detail, multiple characters, and more bizarre events. Stable individual traits, including attitudes toward dreaming, propensity for mind-wandering, and subjective sleep quality, selectively shaped dream content.

Something quietly remarkable is happening to the craft of writing software. In the space of a few years, AI has moved from suggesting the next word of code to drafting whole programs, and now to acting as an autonomous collaborator that can plan, build, and revise with only light human oversight. For researchers and the engineers who support them, this is more than a new tool in the box — it is an invitation to reconsider what it means to build research software at all. If a machine can handle much of the implementation, what is left for the human researcher? The answer, increasingly, lies further upstream: in asking the right questions, framing problems clearly, and exercising the judgement that decides whether an answer is trustworthy. This talk reflects on that shift. We will trace the short but striking journey from autocomplete to autonomous agents, consider how the act of "programming" is being redefined as a conversation between human intent and machine execution, and confront the harder questions this raises — about trust, reproducibility, security, and the future of expertise. The aim is less to instruct than to provoke: to ask what kind of researchers, and what kind of research, this new way of working will produce.