FAIRS is a selective, two-day workshop designed to bring together prominent PhD students and postdoctoral researchers working on foundations of artificial intelligence and machine learning. Relevant areas include: learning theory; robustness, fairness, and privacy; multiagent systems, RL & game theory; generative AI & alignment; causality & causal inference; quantum computing; stochastic processes & diffusion; physics-informed ML.
Aimed at early-career researchers who are preparing to enter the academic and industry job markets, FAIRS will be a platform for participants to showcase their work, exchange innovative ideas, and network with senior leaders in the AI community. Over the course of two days, attendees will give talks about their research, attend panels on navigating the job market, and build ongoing professional collaborations.
Applications for FAIRS 2026 are now closed.
Selected Speakers
We are delighted to welcome the following ten rising stars to FAIRS@Warwick 2026.
Bio: Anay Mehrotra is a Motwani Postdoctoral Fellow at Stanford University, hosted by Amin Saberi. He received his Ph.D. from Yale University, advised by Amin Karbasi and Manolis Zampetakis. His research studies learning with missing and selectively observed data and the foundations and safety of generative AI. His work received the COLT 2025 Best Paper Award, has been featured in WIRED, and has been used to safety-test frontier language models.
Learning Theory in the Wild: From Missing Data to Test-Time Reasoning
What can be learned from data? Traditional answers to this question assume an idealized data-generating process where training and test distributions are identical, which is rarely the case in applications. In the first part of the talk, we will revisit this question for positive-only learning, a setting where only positive examples are observed. This is a challenging problem arising in bioinformatics and causal inference. Classical results show that learning here is impossible in general. However, we will show that the hard instances are fragile: under a smoothed analysis, which rules out pathological distributions, efficient learning is possible. The same ideas also lead to faster and more general algorithms learning from truncated data, a foundational problem in statistics.
In the second part of the talk, we will consider the complementary question: How can we make better use of what a model has learned? In particular, can a language model’s performance be improved at test time without changing its parameters (as in post-training) or increasing its sampling budget (as in test-time scaling)? We will show how to allocate this budget more effectively by choosing which decisions to revisit in a model’s reasoning, rather than simply continuing to sample the next token. We will explain from first-principles why this approach can help and demonstrate improvements on reasoning tasks.
Bio: Ekaterina Kochetkova is a third year CS PhD student at EPFL working with Michael Kapralov. She is broadly interested in theoretical computer science. One direction of current work is applying spectral graph theory to sublinear algorithm design. Another branch is optimization the memory/runtime of LLM inference, emphasizing algorithms with strong theoretical guarantees.
Robustifying Spectral Methods: Combining Graph Structure and Features with Provable Guarantees
Graph learning uses node features and edges representing relationships such as citations or social interactions. How much can we provably gain by combining these two sources? In this talk, I will present an algorithm that robustifies spectral clustering using noisy feature predictions. We consider graphs with well-connected communities separated by sparse cuts, and feature predictions corrupted independently across vertices. When graph structure and feature predictions have error scales $\epsilon$ and $\delta$, respectively, combining them yields misclassification rate $\approx O(\epsilon\delta)$. The algorithm exploits structural properties of vertices with misleading spectral embeddings and classifies individual nodes in sublinear time.
Bio: Elies Gil-Fuster is a last-year PhD student working on quantum computing. His advisors are Jens Eisert and Vedran Dunjko. He has studied different foundational aspects of quantum machine learning models, like their generalization, expressivity, and the potential for quantum advantage. Beyond learning, he is interested in quantum complexity theory, quantum algorithmic techniques, and query complexity. Concretely, he is excited to work on theory questions that capture real-life scenarios.
The quest for good QML models
Machine learning is often cited as a promising direction for applications of quantum computers. Here, one motivating question is “in which situations can we expect a benefit from using quantum learning models instead of classical neural networks?”. At the moment, answering this question relies mostly on theoretical analyses, as the currently available quantum hardware platforms are still small and faulty. In this talk, I will introduce quantum machine learning (QML) for supervised learning on classical data and discuss recent theoretical results on the fundamental properties of quantum learning models, including expressivity, trainability, generalization, and the prospects for quantum advantage. I will conclude by highlighting open problems and promising directions for future research.
Bio: Francesco Quinzan is a Senior Research Associate at the University of Oxford and an Affiliated Senior Research Associate at the University of Cambridge. My research focuses on reliable AI, particularly foundation-model post-training, learning under imperfect feedback, and emerging agentic systems. I serve as an Area Chair for NeurIPS 2026 and lead the ELLIS Workshop on LLM Post-Training at NeurIPS 2026. I have supervised more than 20 Ph.D. and M.Sc. students and previously held a postdoctoral position at KTH.
Reliable Post-Training: Foundation Models under Imperfect Feedback
Post-training enables foundation models to align with human preferences, improve reasoning, and adapt to new tasks, but imperfect feedback can introduce unexpected failure modes. I will present recent work on robust alignment and reinforcement learning for LLM reasoning, and argue for viewing post-training as a sequential, closed-loop learning problem. I will then discuss how these challenges extend to agentic AI, where feedback errors can propagate across interactions and safety failures may emerge over time.
Bio: Gezheng Xu is a Postdoctoral Research Fellow in the Department of Computer Science at Western University, where she earned her PhD. Her research focuses on trustworthy machine learning, with an emphasis on robustness, fairness, and adaptation under distribution shift. She develops reliable learning methods that generalize across domains and actively explores their applications to scientific discovery. Her work has appeared in leading venues, including ICLR, ICML, NeurIPS, and IEEE TPAMI.
Trustworthy Machine Learning under Distribution Shifts: Adaptation, Uncertainty, and Fairness
Recent advances in machine learning and artificial intelligence have achieved remarkable success, yet their reliability often depends on assumptions that may not hold in real-world deployment. Distribution shifts between training and deployment environments, together with heterogeneity across population subgroups, can compromise model performance and fairness. Limited access to source data, unreliable supervision, and uncertainty in model predictions further complicate learning in these settings, highlighting the need for principled approaches to trustworthy machine learning.
This talk examines two key dimensions of trustworthy machine learning under distribution shifts: adaptation across domains and fairness across population subgroups. It explores how understanding pseudo-label noise and controlling uncertainty can improve source-free domain adaptation, spanning both classification and recent advances in regression. It also addresses fairness challenges arising from limited or biased data, with an emphasis on learning across multiple subgroups and discovering intersectional populations whose risks may be obscured by aggregate performance. Together, these investigations connect the technical and societal aspects of reliable learning, contributing toward AI systems that adapt more reliably to changing environments and serve diverse populations more equitably.
Bio: Kavya Ravichandran is a postdoctoral scholar at Stanford University, working with Sanmi Koyejo and Rob Reich in the STAIR and AI Governance Labs. Previously, she completed her PhD at Toyota Technological Institute at Chicago (TTIC), advised by Avrim Blum, and her MEng and S.B. at MIT. She works on algorithms for decision-making, steering AI use toward positive social outcomes, and connections between theoretical computer science and philosophy. Outside of research, Kavya enjoys getting young students excited by math and computer science.
Sequential and Social Decision-Making: Improving Multi-Armed Bandits and Pessimism Traps
We are often faced with decisions that require us to invest effort before reaping reward. For instance, persevering through higher education or starting a business. In this talk, I will share my recent work on theoretical models for making such decisions based on (1) one’s interaction with the environment and (2) decisions of others in one’s community. First, I will discuss the improving multi-armed bandits problem, in which an agent is deciding between k different options, each of which has more reward the more they interact with it. The agent’s goal is to maximize their total reward. We develop near-optimal algorithms for solving this problem in the general case. Next, I will describe a work on how to escape pessimism traps, a phenomenon in which agents are influenced by their predecessors to engage in less-ambitious goals. We provide a mathematical formalism in which to study this concept that was first identified by philosophers. Then, we provide an algorithmic intervention to sustainable shift communities out of these traps. I will conclude the talk with some reflections on how these two styles of work can influence each other and my plans for future work in these areas.
Bio: Max Springer is a postdoctoral research fellow at Princeton University, where he studies the algorithms and AI systems that increasingly make consequential decisions in elections, hiring, and public institutions. His work spans designing algorithms with provable fairness and efficiency guarantees, developing methods to measure and verify the alignment of large language models, and stress-testing public infrastructure against increasingly capable AI. He is also a scientific consultant to the United Nations Independent International Scientific Panel on AI, where he helped draft its Preliminary Report. Springer received his Ph.D. in Applied Mathematics from the University of Maryland, College Park, advised by MohammadTaghi Hajiaghayi and supported by an NSF Graduate Research Fellowship.
From Ballots to Bridges: Formalizing Algorithms as Public Infrastructure
Building safe and reliable AI systems requires a shared, precise formalism for what we expect of them. Historically, durable regulation of a technology has followed such formalisms, as risk-limiting audits did for elections and differential privacy did for the U.S. Census. In this talk, I illustrate my research agenda of formalizing, measuring, and institutionalizing guarantees for algorithmic decision-making in the context of election systems. I use this illustration to identify the chief responsibility of theoretical researchers now that AI models increasingly prove theorems themselves: deciding what to formalize. I then show how I am putting these ideas into practice to better protect and regulate algorithmic public infrastructure.
Bio: Oxana Shaya is currently a PhD candidate in quantum information theory at Leibniz University Hannover. She holds a master’s degree in physics from RWTH Aachen University. Her research asks what makes quantum systems complex, when they are hard to learn, and how their symmetries can be identified. Her broader interests lie in quantum foundations and identifying the physical resources that can enable universal function approximation in quantum learning models.
Fundamental Barriers to Quantum Generative Learning
The output distributions of parameterized quantum circuits can serve as generative models for classical data. Such models, known as quantum circuit Born machines, are one approach to quantum generative modeling, motivated by potential computational advantages in sampling. In this talk, I examine concentration of measure as a source of fundamental barriers to distribution learning and model training that even persist under symmetry restrictions. I present average-case hardness results in the statistical-query framework and discuss their consequences for training quantum generative models. I also show how loss estimation and loss design affect Born-machine trainability.
Bio: Sina Akbari is a postdoctoral researcher at the Statistical Laboratory, University of Cambridge. He received his PhD in Computer Science from EPFL. His research focuses on causal inference, with a focus on developing methods that combine flexible machine learning with rigorous statistical guarantees. His interests include causal identification and discovery, experiment design, partial identification, semiparametric statistics, and sequential decision-making. More broadly, he is interested in making reliable causal conclusions possible in complex, data-rich settings where classical assumptions are too restrictive.
From Non-Identification to Point or Partial Identification: Designing the Right Experiments
Many questions about the behaviour and impact of complex systems are causal: what would happen if we changed a component of an established pipeline? Why did an AI model act differently in two similar contexts? Such questions may not be answerable by collecting more data alone. On the other hand, randomising the variable of interest can be infeasible, too costly or even unethical. What experiments should we conduct instead and what can we learn from these ‘proxy’ experiments?
This talk follows the path from non-identification to point or partial identification. I will first introduce the essential ideas of causal identification using graphical models. I will then show how experiments on carefully chosen proxy variables can identify effects that cannot otherwise be recovered and formulate the choice of experiments as an optimisation problem. This reveals connections with familiar computational problems including minimum hitting sets, satisfiability, and integer programming, leading to more scalable algorithms. Finally I will discuss what can still be learned when an exact causal effect cannot be identified using instrumental-variable bounds as an example. A running example involving retrieval-augmented AI systems will illustrate the main ideas.
Bio: Yongqiang (Andrew) Chen is a postdoctoral researcher at MBZUAI and CMU. He received his Ph.D. from CUHK in 2024. His research focuses on reliable learning, reasoning, and discovery in the open world, spanning out-of-distribution generalization, long-horizon reasoning, and LLM agents for scientific and algorithmic discovery. His work has appeared at ICML, NeurIPS, and ICLR, including oral and spotlight presentations.
Toward Reliable Long-Horizon Reasoning for AI Scientist Agents
AI scientist agents are a north-star application of large language models. Reliable discovery requires them to reason over long horizons: infer mechanisms from imperfect observations, revise their beliefs as evidence accumulates, and retain reusable ideas that guide subsequent experiments.
In this talk, I examine three fundamental capabilities required for building reliable AI scientists. First, through an interactive benchmark, CausalGame, that imitates real-world scientific discovery, I show that existing frontier LLMs can hardly form proper beliefs from experimental observations and often confuse correlation with causation. Second, in long-horizon reasoning, imperfect belief updates can trap agents, where they stop taking informative actions to make progress. I develop T3, a principled method that restores proper credits promoting informative actions and belief updates. Finally, I present CausalEvolve, which extracts reusable high-level factors from evaluated programs, maintains them in structured memory, and uses them to guide evolutionary search. This high-level guidance improves both sample efficiency and final performance across a range of discovery tasks.
Together, these works point toward long-running agents that interpret evidence reliably, update their beliefs, and accumulate reusable knowledge. I close with directions for transferring experience across tasks and horizons for accelerated and continual discovery.
