Research Seminars 2026 – 2027

Title: The Perverse Economics of Generative AI

Abstract: Generative AI has become one of the largest economic undertakings in modern history, with more than $700 billion in AI investment projected for 2026 alone. That capital is a bet not only on what these systems can do, but on the markets forming around them, and on who trains them, who pays for access, who competes, and who is displaced. In this talk, we show theoretically and empirically that the dominant pricing mechanism for large language models—charging a fixed price per token—creates a perverse financial incentive for self-serving providers to strategically manipulate different aspects of the generative process at the expense of the users. Finally, we discuss several solutions to eliminate this misalignment, including alternative pricing mechanisms and third-party auditing.

Bio: Manuel Gomez Rodriguez is a tenured faculty member at the Max Planck Institute for Software Systems (MPI-SWS). His research interests lie in the development of human-centric machine learning models and algorithms. He has co-authored over 50 publications in top-tier conferences (NeurIPS, ICML, UAI, AISTATS, AAAI, KDD, WWW) and journals (PNAS, Nature Communications, Management Science, JMLR, TMLR). Manuel is an ELLIS Fellow and has received several recognitions for his research, including an ERC Consolidator Grant, an ERC Starting Grant, an outstanding Paper Award at NeurIPS and a Best Research Paper Honorable Mention at KDD and WWW. Manuel holds a M.S. and a Ph.D. in Electrical Engineering from Stanford University (2009 and 2013) and a B.S. in Electrical Engineering from University Carlos III in Madrid, Spain (2006).

Title: What Makes Local Updates Effective? A Fine-Grained Theory of Local SGD under Second-Order Heterogeneity

Abstract: Local SGD, also known as Federated Averaging, is a simple approach to communication-efficient distributed optimization in which each client performs several stochastic-gradient steps between communication rounds. Despite its empirical success [1], a basic theoretical question remains: when do sequential local updates outperform the use of the same stochastic gradients in a large mini-batch? In this talk, I will summarize a line of work that develops a hierarchy of assumptions on data heterogeneity and clarifies when local updates are worthwhile [1, 2, 3, 4, 5, 6, 7, 8].

We begin with the homogeneous setting, reviewing the optimal rates and thereby establishing the benchmark that any heterogeneous guarantee could hope to match [2, 3]. We then study a natural first-order condition capturing approximate simultaneous realizability and show that it is too weak to guarantee an advantage for Local SGD [4]. At the opposite extreme, uniform gradient-similarity assumptions can recover the improvements seen in the homogeneous setting [5], but at the cost of excluding meaningful heterogeneity across clients [4]. These observations, together with analyses of related local-update methods in nonconvex optimization [6], motivate bounded second-order heterogeneity, which controls differences in local objective geometry rather than requiring local gradients to remain uniformly close. I will review the resulting theory in the strongly convex setting [7], including convergence rates and the roles of heterogeneity, stochastic noise, and higher-order smoothness, before presenting recent guarantees for general convex objectives and nearly matching lower bounds [8]. A central idea in these analyses is a trajectory-dependent treatment of client drift: consensus error is controlled using only the gradient disagreement encountered along the optimization path, and the resulting bounds are closed through self-bounding recursions. The theory identifies regimes in which Local SGD provably outperforms mini-batch SGD and suggests a simple principle: effective local updates require aligned local geometry rather than globally similar gradients. I will conclude by discussing how these insights and techniques inform our understanding of data heterogeneity in other distributed optimization problems.

Bio: Kumar Kshitij Patel is a Postdoctoral Associate at the Yale Institute for Foundations of Data Science. During Spring 2026, he was a Fellow at the Simons Institute for the Theory of Computing at the University of California, Berkeley. His research focuses on the foundations of collaborative and federated learning, with an emphasis on distributed optimization, data heterogeneity, differential privacy, and client incentives. Kshitijreceived his PhD in 2025 from the Toyota Technological Institute at Chicago, where he worked with Prof. Nathan Srebro and Prof. Lingxiao Wang. He completed his undergraduate degree in computer science at the Indian Institute of Technology, Kanpur. He is a recipient of the IJCAI 2024 Distinguished Paper Award.

Title: Particle-based Generalised Stochastic Optimisation

Abstract: We develop a class of general diffusion-based stochastic particle optimisation methods for loss functions with intractable gradients. Specifically, we consider problems in which the loss gradient is an integral with respect to a parameter-dependent distribution, a structure that includes training generative models, fine-tuning, and learning latent-variable models. We introduce mean-field dynamics and its interacting-particle approximations, which contain several existing algorithms as special cases and provide a route for constructing new methods. Under well-posedness and joint contractivity assumptions, we prove exponential convergence and show that the continuous-time particle system admits a non-asymptotic error bound. We illustrate our framework by developing higher-order Langevin and momentum optimisers and evaluating them on maximum marginal-likelihood estimation and energy-based-model training.

Joint work with Jiechen Jackie Zhang

Bio: Deniz Akyildiz is an Associate Professor in Statistics in the Department of Mathematics at Imperial College London. His research lies at the intersection of computational statistics and machine learning, with a focus on sampling and Monte Carlo methods. This includes interacting particle algorithms for learning problems, Bayesian filtering for state-space models, as well as diffusion-based methods for sampling, inference, and inverse problems. He has published in leading machine learning conferences (NeurIPS, ICML, AISTATS, UAI) and journals (JMLR, Bernoulli, Statistics and Computing, SIAM Journal on Numerical Analysis, SIAM/ASA Journal on Uncertainty Quantification). He is an Associate Editor for Statistics and Computing and a member of the Royal Statistical Society's Computational Statistics and Machine Learning Section Committee. Deniz holds a BSc and MSc in Electronics and Communications Engineering from Istanbul Technical University and a PhD from Universidad Carlos III de Madrid. He was a postdoc at the Department of Computer Science in Warwick and The Alan Turing Institute before joining Imperial.

Title: Beyond Spillovers: Disentangling Causal Effects on and Through Social Networks

Abstract: Interventions in connected populations often exhibit complex dynamics, where effects travel along existing social ties, alter the ties themselves, and propagate through these network changes. This talk presents weighting estimators to evaluate causal effects under static and dynamic network interference by conceptualizing the network across three distinct roles: as a channel, an outcome, and a mediator.

First, to understand how far treatment effects travel through the network as a channel, we propose a framework for estimating higher-order spillover effects under partial interference. By defining distance-specific hypothetical assignments, we develop Horvitz-Thompson, Hájek, and weighted least squares estimators that do not require an analyst-specified exposure mapping, thereby protecting against common misspecification biases. Second, recognizing that interventions can actively rewire social connections and act as an outcome, we construct a dyad-level causal framework to evaluate effects on the ties themselves. This approach defines direct, spillover, and total effects to quantify how interventions impact the formation of new ties and the retention or dissolution of existing ones.

Finally, bridging traditional interference and mediation analyses, we conceptualize dyadic ties as mediators and define mediated and non-mediated direct, spillover, and ego-alter effects under a hypothetical Bernoulli treatment allocation. For observational clustered data, we construct multiply robust estimators that combine an outcome model, a dyadic mediator model, and a cluster-level propensity score, achieving consistency whenever any two of the three models are correctly specified. These methodologies are validated through simulation studies and applied to two empirical settings. Using data from a maternal and child health trial in Honduras, we evaluate information diffusion and program-induced network rewiring. Additionally, applying our mediation framework to observational data from 75 villages in Karnataka, India, we quantify the extent to which the dissolution of informal ties mediated the effects of microfinance exposure on household borrowing.

Short Bio: Laura Forastiere, PhD, is an Associate Professor in the Department of Biostatistics at Yale University, with a secondary appointment in the Department of Statistics and Data Science.

Dr. Forastiere’s academic and methodological work operates at the cutting edge of causal inference, social network analysis, and policy evaluation. She specializes in developing and implementing advanced statistical frameworks tailored for complex clustered and network structures—environments where traditional statistical theories frequently break down. Her methodology maps the intricate ways interventions ripple through communities, evaluating key dynamics like peer influence, spillover effects, and causal mechanisms.

In the fields of global health and development economics, Dr. Forastiere applies rigorous methods to evaluate programs for vulnerable populations in low- and middle-income countries, including Honduras, Brazil, India, Uganda, Rwanda, and South Africa. Her extensive field research explores the socioeconomic and health impacts of structural initiatives, including health insurance policies, microcredit programs, conditional cash transfers, and nutritional interventions. Additionally, her diverse portfolio spans large-scale program evaluations addressing maternal and child health, cognitive development, and infectious diseases such as malaria and HIV. Driven by a mission to reduce global inequalities, Dr. Forastiere focuses on designing innovative, cost-effective, and sustainable strategies that leverage social interactions, cooperative behaviors, and community engagement. Prior to her tenure track at Yale, she conducted postdoctoral research within the Department of Statistics at Harvard University and the Human Nature Lab at the Yale Institute for Network Science.

Title: Learning the score under shape constraints

Abstract: Score estimation has recently emerged as a key modern statistical challenge, due to its pivotal role in generative modelling via diffusion models. Moreover, it is an essential ingredient in a new approach to linear regression via convex M-estimation, where the corresponding error densities are projected onto the log-concave class. Motivated by these applications, I will discuss some recent results on the minimax risk of score estimation for log-concave densities on the real line. [Based on this arXiv paper.]

Short Bio: Rebecca Lewis is a research fellow in the Department of Statistics at the University of Oxford. Prior to this, she completed her PhD at Imperial College London under the supervision of Heather Battey. Rebecca’s research lies in statistical theory and methodology, with a particular focus on shape-constrained and high-dimensional inference.

Title: Towards Trustworthy Cooperative AI: Simulating, Aligning, and Governing Multi-Agent Systems

Abstract: As AI systems increasingly operate as autonomous agents, intelligence is becoming inherently social. Future AI systems will not act in isolation, but will learn, coordinate, communicate, and make decisions alongside humans and other agents in open-ended environments. This shift raises fundamental challenges: How can agents acquire cooperative behaviours? How can they adapt to human preferences and evolving social norms? And how can we ensure safety when intelligent agents interact at scale?

In this talk, we present a framework to ground cooperative AI in real-world human contexts. We will explore how to resolve the computational limits of complex social simulation, transition from hand-crafted rewards to dynamic value alignment via mixed-quality human feedback, and replace inflexible numerical costs with intuitive natural language safety boundaries. Together, these interconnected works outtline a cohesive pathway for building cooperative AI that is highly scalable and deeply aligned with human intent. Finally, I will outline a research agenda for building trustworthy agent societies where learning, coordination and safety continue beyond deployment.

Bio: Dr Yali Du is an Associate Professor in AI at King’s College London and a Turing Fellow at The Alan Turing Institute. She leads the Cooperative AI Lab. Her research aims to build cooperative, safe, and human-aligned AI agents for complex decision-making tasks, with a focus on multi-agent cooperation, human-AI coordination, reinforcement learning, and value alignment. She received the AAAI New Faculty Highlights Award in 2023 and was named a Rising Star in AI (KAUST). She received Responsible AI UK Best Research Paper Awards in 2026. She has delivered tutorials on cooperative multi-agent learning at ACML 2022 and AAAI 2023. She serves as an Associate Editor for the Journal of AAMAS and IEEE Transactions on AI, and has served as an Area Chair for NeurIPS/ICML/ICLR/IJCAI and as a Senior Area Chair for AAMAS 2026-2027, NACCL2027. She has also served on the organising committees of DAI 2025, NeurIPS 2024 and AAMAS 2023&2027.