Young Researchers Workshop 2026

Poster Session One (Thursday, October 8, 8:30 a.m. – 9:30 a.m. in Duffield Atrium)

Poster Session Two (Friday, October 9, Noon – 1 p.m. in Duffield Atrium)

Information for Speakers

All talks will take place in 230 Anabel Taylor Hall. The room has built in projectors, microphones and is zoom capable. We hope to record your talk; please let us know if that is not possible. An in-room computer can be used to present from pdf and powerpoint files. You also have the option of presenting from your own laptop using zoom. Speakers have 25 minutes each. We suggest using 20 minutes for the talk and leaving 5 minutes for questions.

Information for Poster Presenters

Poster sessions are planned for 8:30 a.m. on Thursday and noon on Friday. If you are accepted for a poster presentation, then you will be allocated to one or the other of these sessions. We will provide posterboards and pushpins. There will be enough space at least for 36×48 inch posters, and probably for larger posters if you prefer. You can print your poster at the Mann Library: prices and hours of operation.

Talk Abstracts

Optimization Algorithms – Technical Session 1

Liviu Aolaritei, EECS, UC Berkeley
Beyond Optimal Rates in Stochastic Optimization: Trajectory-Adaptive Stopping Rules

Stochastic gradient descent (SGD) is typically analyzed at a deterministic horizon chosen before the algorithm is run, even though practical stopping decisions are made adaptively by inspecting the evolving trajectory. This mismatch creates a fundamental certification problem: fixed-time guarantees do not generally remain valid at data-dependent stopping times, while deterministic horizons derived from worst-case bounds can be highly conservative. We address this problem for strongly convex stochastic optimization by constructing fully observable, trajectory-adaptive upper confidence sequences for the squared distance of the last iterate to the optimizer and the suboptimality of a weighted average. These bounds hold simultaneously over time, attain the optimal 1/t decay rate up to iterated-logarithmic factors in the worst case, and adapt to the realized stochastic gradients, allowing SGD to stop as soon as a prescribed accuracy is certified without sacrificing statistical validity.

Our approach treats the evolving SGD trajectory as a sequential experiment whose observations provide evidence about the unknown optimization error. To formalize this perspective, we develop new recursive confidence-sequence techniques and a general time-uniform empirical Bernstein inequality for adapted processes with time-varying conditional means and predictable ranges that may grow without bound. We further extend these confidence-sequence constructions to minibatch SGD, with the empirical Bernstein bounds exploiting the realized second-moment structure within each minibatch. Numerical experiments show that the resulting stopping rules can require several orders of magnitude fewer iterations than natural deterministic horizons.

Wenzhi Gao – ICME, Standford University
Gradient Methods with Online Scaling

We introduce a framework to accelerate the convergence of gradient-based methods with online learning. The framework learns to update the stepsize in gradient descent with online learning and provably accelerates gradient-based methods. A key insight is to view stepsize selection in gradient descent as a sequential decision making problem. We also discuss implications of the framework, including convergence properties, extensions, as well as its connection to the quasi-Newton methods. Numerical experiments on deterministic convex and nonconvex problems demonstrate the promising performance of our method.

Ariel Goodwin – CAM, Cornell University
Convex Optimization in Geodesic Spaces: Problems and Algorithms

In Euclidean space, convexity lies at the heart of designing practical and theoretically sound algorithms for a variety of problems in optimization. But there are several applications that are naturally modeled as optimization problems on non-Euclidean metric spaces, ranging from manifolds to trees. In this talk we present a survey of recent work that situates many of these application domains within a common framework: optimization on geodesic spaces with curvature bounded above. Such nonsmooth and nonlinear spaces have much in common with Euclidean space when it comes to convexity, but they also exhibit many subtle differences. In particular, the lack of linear structure in such spaces makes it challenging to define a suitable notion of subgradient. We will present a new concept which we call Busemann subgradients, based on a nonlinear generalization of affine functions. These Busemann subgradients give rise to implementable Busemann subgradient algorithms with complexity matching their Euclidean predecessors. We illustrate these algorithms on fundamental optimization problems such as computing weighted averages and minimum enclosing balls.

Learning for Decision-Making – Technical Session 2

Yifan Wu – Microsoft Research, New England         
Calibration Error for Decision Making

In modern AI workflows where prediction and decision-making are separated, an upstream prediction team must serve unknown or heterogeneous downstream decision makers. Calibration, a property of predictive models, enables this separation. It requires reported probabilities to be conditionally unbiased, so they can be trusted and directly used by any downstream decision maker without task-specific correction.

We propose Calibration Decision Loss (CDL) as a decision-theoretic calibration error. CDL is defined as the largest improvement in expected payoff obtainable by recalibrating a predictor across all bounded decision problems. By definition, CDL captures the worst-case decision loss from miscalibration and therefore directly measures whether a predictive model is trustworthy for heterogeneous downstream users.

Our results highlight a fundamental difference between decision-theoretic error and known calibration errors. CDL counts only biases that can change decisions, whereas standard metrics such as Expected Calibration Error (ECE) average over all prediction biases, regardless of their impact on decisions. In an online setting, this perspective leads to a prediction algorithm with near-optimal expected CDL, improving upon decision-making guarantees obtained by optimizing canonical calibration errors.

Shang Liu, Imperial Business School, Imperial College London
Learning from Learned Feedback

Artificial intelligence advances by learning from existing examples and by improving through trial and error. Human text and expert demonstrations provide valuable starting points; reliable, inexpensive feedback allows learning to continue beyond those examples. Yet in many applications, generating new behavior is much easier than evaluating it. Human judgments are costly, physical experiments are limited, and the long-term consequences of an action may be difficult to assess.

A common response is to learn the feedback itself. For example, in RLHF, a reward model learns to approximate human preferences, while in RL a critic learns to estimate the future return of an action. These learned evaluators turn limited experience into feedback that can guide further improvement. But they also introduce a fundamental tension: a learner can become better at satisfying its evaluator without becoming better at the underlying task. The need for robustness therefore arises within the learning process itself, even when the external environment does not change.

In this talk, we present distributionally robust regret optimization (DRRO) as a unified approach to this problem. The central idea is to account for uncertainty in the learned evaluator by asking how much a decision could lose relative to the best decision under each plausible evaluation. This provides a way to use the information in a proxy while limiting the consequences of trusting it too strongly. Through applications to reinforcement learning from human feedback and actor-critic reinforcement learning, we show how this principle leads to practical learning algorithms and improved performance in examples such as the language-model alignment and the continuous control.

Matthew Zurek, Computer Sciences, UW-Madison
When Can We Learn Average-Reward Policies from Offline Data?

This talk considers average-reward reinforcement learning in the offline setting, where the goal is to learn a policy that maximizes the long-term rate of reward using only a fixed existing dataset. I will present an algorithm whose data coverage requirements depend on the properties of a single good policy, rather than worst-case properties across all possible policies. A central finding is that data covering a policy’s long-run behavior is insufficient: learning also provably requires observations from states that the policy visits only transiently. Surprisingly, the amount of data needed from these transient states can remain essentially fixed even as we seek increasingly accurate policies. The algorithm is minimax-optimal and does not need to know the relevant properties of the good policy in advance; this adaptivity is achieved through a sharp penalty function based on a quantile-clipped estimator.

Finance and Large Stochastic Systems – Technical Session 3

Yinbin Han, MS&E, Stanford University
Generative Diffusion Models for Complex Optimization and Decision-Making

Diffusion models offer a new class of data-driven methods for high-dimensional stochastic optimization and decision problems. I will first discuss score matching and statistical guarantees for gradient-based estimation of diffusion models. I will then introduce adaptive generation for sequential problems with information flow constraints. As a financial application, I will leverage this framework to model implied volatility surfaces simulation, and show how a fine-tuning approach can reduce violations of no-arbitrage conditions. Finally, I will present a data-driven hedging problem that uses the learned model to construct decisions under market uncertainty.

Qinxin Yan, IEOR, UC Berkeley
Learning When to Stop: Stopping Games and Learning Approximate Equilibria

When should a firm enter a market, knowing that waiting preserves flexibility but competitors may invest first? Such decisions lead to optimal stopping games, where each player’s payoff depends on both an uncertain environment and others’ stopping decisions.

This talk connects these stopping games to stochastic control and reinforcement learning. We represent randomized stopping decisions through cumulative stopping probabilities and use an approximate potential framework to reduce strategic interaction to a single optimization problem. The resulting singular-control formulation admits an HJB system that captures waiting, gradual stopping, and finite jumps. We establish a viscosity characterization, regularity, and uniqueness under suitable assumptions.

Building on this control formulation, we develop an actor–critic algorithm that learns stopping policies from state transitions and rewards. Numerical experiments compare learned policies with an analytical benchmark and assess approximate equilibrium through unilateral deviations. Based on joint work with Xin Guo and Mehdi Talbi.

Xinyu Li, Mathematical Institute, University of Oxford
An α-Potential Game Framework for N-Player Dynamic Games

Motivated by game-theoretic models of crowd motion dynamics, this talk analyzes a broad class of distributed games with jump diffusions within the recently developed α-potential game framework. Analogous to static potential games (corresponding to α = 0), this framework reduces the challenging task of finding α-Nash equilibria for an N-player dynamic game to minimizing an α-potential function.

We show that analyzing α-Nash equilibria reduces to solving a finite-dimensional control problem. Beyond viscosity and verification characterizations for general games, we examine how spatial population distributions and interaction rules shape the structure of α-Nash equilibria. For crowd motion network games, we show that α = 0 for all symmetric interaction networks; for asymmetric networks, we quantify the polynomial and logarithmic decay of α in terms of the number of players, network degree, and decay rate of interaction asymmetry. Our analysis accommodates heterogeneous preference parameters, extending beyond the mean-field interactions commonly considered in the literature.

Finally, numerical implementations using policy gradient-based algorithms demonstrate the computational advantages of the α-potential game framework for computing Nash equilibria in general dynamic games

Online Algorithms and Allocation – Technical Session 4

Anders Wikum, MS&E, Stanford University
Algorithms with Calibrated Machine Learning Predictions

The field of algorithms with predictions incorporates machine learning advice in the design of online algorithms to improve real-world performance. A central consideration is the extent to which predictions can be trusted — while existing approaches often require users to specify an aggregate trust level, modern machine learning models can provide estimates of prediction-level uncertainty. We propose calibration as a principled and practical tool to bridge this gap, demonstrating the benefits of calibrated advice through two case studies: the ski rental and online job scheduling problems. For ski rental, we design an algorithm that achieves near-optimal prediction-dependent performance and prove that, in high-variance settings, calibrated advice offers more effective guidance than alternative methods for uncertainty quantification. For job scheduling, we demonstrate that using a calibrated predictor leads to significant performance improvements over existing methods. Evaluations on real-world data validate our theoretical findings, highlighting the practical impact of calibration for algorithms with predictions.

Pranav Nuti, Booth School of Business, University of Chicago
Stationary Online Contention Resolution Schemes and Applications to Reusable Resource Allocation

In this talk, motivated by problems in Bayesian reusable resource allocation, I will introduce the concept of stationary online contention resolution schemes (S-OCRSs). OCRSs are a central tool used to solve non-reusable resource allocation problems. They convert solutions to fluid approximations of problems into feasible online policies while approximately preserving allocation probabilities. The defining feature of S-OCRSs is that they ensure that the probability of allocating any given set of products is independent of the arrival order of requests. I will show how S-OCRSs can be used to solve reusable resource allocation problems, and discuss a general ‘maximum-entropy’ approach to construct and analyze S-OCRSs.

Our approach, using a unified method for a variety of feasibility constraints, obtains results that match the state-of-the-art in the non-reusable case, and even improves it for a bipartite matching feasibility constraint. Our results also extend from the resource allocation setting to assortment optimization setting, and our policies can be implemented using prices.

Siyue Liu ACO, Carnegie Mellon University
One Good Away: Fair Allocation under Combinatorial Constraints

Fair division of indivisible items has attracted substantial attention over the past decade. Without feasibility constraints, an allocation that is envy-free up to one good (EF1) always exists for agents with monotone valuations and can be found efficiently. In many real-world scenarios, however, each agent’s bundle must also satisfy operational requirements, such as capacity or partition constraints. Matroids provide a unified framework for modeling such requirements. We prove that, under a common matroid constraint, an EF1 allocation always exists when agents have identical additive valuations. We also give a strongly polynomial-time algorithm to compute such an allocation. Our result builds on two advances in matroid theory: resolving the matroid equitability conjecture, which concerns balancing a designated subset of items across feasible bundles, and its weighted analogue, which enables the balancing of arbitrary item values.

Games, Equilibria and Optimization – Technical Session 5

Jingye Xu, ISYE, Georgia Institute of Technology
Asymptotically Strong Dual of Smooth Nonconvex Problems: A Smooth Shapley–Folkman Perspective

The Shapley–Folkman Lemma, a foundational result in convex geometry, asserts that the nonconvexity of a Minkowski sum of n-dimensional bounded nonconvex sets does not accumulate once the number of summands exceeds the dimension n, and thus the sum becomes approximately convex. This Lemma, originally published by Starr is used to show quasi-equilibrium in nonconvex market models in economics and equivalently used to estimate the duality gap of the Lagrangian dual of separable nonconvex problems. Given its foundational nature, we pose the following geometric question: Is it possible for the nonconvexity of the Minkowski sum of n-dimensional nonconvex sets to vanish asymptotically, rather than merely converge to a constant, as the number of summands increases?

We answer this question affirmatively. First, we provide an elementary geometric proof of the Shapley–Folkman Lemma based on the facial structure of the convex hull of each set under a pointedness assumption. This leads to a small refinement on the classical error bound derived from the Lemma in the pointed case. Building on this new geometric perspective, we next show that if the sets satisfy a certain local smoothness condition, which naturally arises, for example, in the epigraphs of smooth functions, then their Minkowski sums converge directly to a convex set as the nonconvexity measure vanishes asymptotically.

This result has direct implications for optimization and economic theory. In particular, we show that the Lagrangian dual of block-structured smooth nonconvex problems, possibly subject to additional sparsity constraints, is asymptotically tight under mild assumptions. This sharply improves upon classical Shapley–Folkman based bounds, which generally predict a nonvanishing duality gap. Equivalently, from an economic perspective, this result implies that, asymptotically as the size of the economy increases, competitive equilibrium exists with smooth utility functions, notwithstanding underlying nonconvexities.

Nicolas Lanzetti, Caltech
Strategically Robust Decision-Making

In many multi-agent settings, AI agents must make decisions against the uncertain behavior of other AI agents and humans. This uncertainty arises from multiple sources, e.g., incomplete information or bounded rationality. While each source could be modeled individually, their joint presence makes this task daunting. It is thus natural to seek protection against deviations around the emergent behavior itself. In the first part of the talk, we propose that each agent makes decisions in the face of the worst-case behavior contained in an ambiguity set of tunable size, centered at the emergent behavior so implicitly defined, giving rise to a novel equilibrium notion we call strategically robust equilibrium. When judiciously operationalized via optimal transport, strategically robust equilibria (i) interpolate between Nash and security strategies; (ii) come at no additional computational cost compared to Nash equilibria; (iii) often lead to better decisions and higher payoffs. Through experiments in bimatrix games, congestion games, and Cournot competition, we show that strategic robustness protects against uncertainty in the opponents’ behavior and, surprisingly, results in higher equilibrium payoffs for all players, an effect we call coordination-via-robustification. In the second part of the talk, we turn strategic robustness into a scalable training mechanism for AI agents, by introducing a fictitious agent that acts adversarially while staying close to the others’ policies. Across standard multi-agent reinforcement learning environments, warehouse operations, and fine-tuning of LLMs, strategically robust agents generalize to unseen partners, do not free ride, and sometimes achieve higher rewards than standard agents, demonstrating coordination-via-robustification at scale.

Xizhi Tan, Computer Science, Stanford University
Self-Funded Marketplace

We introduce the problem of designing mechanisms that incentivize strategic agents to form self-funded marketplaces. Each agent has a public value, a public revenue and a private cost. The principal seeks to maximize the total value of the selected agents under self-funding, the total payments to the selected agents cannot exceed the revenue they generate. This model captures large-scale platforms, such as content-creator ecosystems and data markets, in which funding arises endogenously rather than being fixed in advance. To evaluate the performance of such mechanisms, we first consider the first-best benchmark (the optimal value achievable in the absence of any private information) and show that no truthful auction can achieve a bounded approximation of this benchmark. Also, even in restricted settings, no auction can achieve better than a logarithmic approximation. We complement these results by proposing a class of sequential auctions whose subgame perfect equilibria guarantee a logarithmic approximation of this benchmark. We then introduce an alternative benchmark, the maximin share (MMS), that better captures the thickness of the market, and we provide an auction whose subgame perfect equilibria achieve a constant approximation of this benchmark.

Operations Research in Practice – Technical Session 6

Matthew Brun, Operations Research Center, MIT
Modeling Adversarial Wildfires for Power Grid Disruption

Electric power infrastructure faces increasing risk of damage and disruption due to wildfire. Operators of power grids in wildfire-prone regions must consider the potential impacts of unpredictable fires. However, traditional wildfire models do not effectively describe worst-case, or even high-impact, fire behavior. To address this issue, we propose a mixed-integer conic program to characterize an adversarial wildfire that targets infrastructure while respecting realistic fire spread dynamics. We design a wind-assisted fire spread set based on the Rothermel fire spread model and propose principled convex relaxations of this set, including a new relaxation of the inner product over Euclidean balls. We present test cases derived from the recent Park, Eaton, and Palisades fires in California and solve models to identify the minimum time-to-outage of multiple-element contingencies and the maximum load shed associated with a sequence of element outages caused by a realistic wildfire. We use the minimum time-to-outage values to screen contingencies and construct security-constrained optimal power flow models that promote operational resilience against wildfire.

Justin Whitehouse – MS&E, Stanford University     
Measuring Gift Card Program Incrementality via Causal Data Fusion

Businesses regularly offer gift card programs to drive customer spending and increase engagement. A central question is how much incremental revenue these programs generate, and which distributional channels drive it most efficiently. Measuring the incremental revenue associated with a gift card program is a challenging problem in causal inference, requiring a firm to infer how much each customer would have spent if they never received a gift card. Observational data on past customer purchasing behavior reveal possession of a gift card only when a customer makes a purchase, thus leaving a customer’s treatment status systematically censored.

In this talk, we develop a novel data fusion approach to overcome this missing data challenge. We identify and estimate incrementality by combining a large observational dataset with a smaller experimental dataset from a different population. Our approach relies on a mild transferability condition, which posits that the conditional relative treatment effect of gift card receipt on the decision to purchase is invariant across the two populations. We develop a flexible, machine learning-based estimator for the incremental revenue and establish its asymptotic normality.

We apply our estimator across both first- and third-party channels through which Airbnb distributes gift cards, finding heterogeneity in incrementality across segments of the population. In particular, we find not only that third-party channels are more incremental than first-party ones but also that “self-gifters” (i.e., customers likely to have purchased their own gift cards) are more incremental than the broader population.

Marouane Ibn Brahim – ORIE, Cornell University   
AI-Mediated Assortment Optimization

AI shopping assistants are increasingly used for product recommendations and influence which products consumers consider. We introduce a model in which an AI intermediary forms a segment-specific shortlist from a retailer’s assortment and customers choose from it according to a multinomial logit model. Under deterministic mediation, the AI displays the highest-ranked available products in a fixed segment-specific ranking. Under stochastic Plackett–Luce (PL) mediation, it displays products with the highest random scores, whose ordering follows the PL model. We construct a ranking dataset by prompting models from OpenAI, Anthropic, and Google with real product listings across retail categories and customer personas, and we estimate PL models on the generated dataset.

We combine these estimates with synthetic customer preferences in experiments on the retailer’s assortment optimization problem. Across seven tested combinations of assortment-size limits and display capacities, assortments optimized as though customers choose directly from the entire assortment incur mean revenue losses of 2.12\%–10.78\% when evaluated under PL mediation, relative to its optimum. Assortments optimized under PL mediation are smaller on average than those optimized without mediation. When evaluated under PL mediation, assortments optimized using fixed rankings from the estimated AI preferences have mean revenue losses below 0.5\% relative to the same optimum in each configuration. Both optimization problems are NP-hard even with one displayed product when the number of market segments is unrestricted. PL mediation remains NP-hard with two market segments and two displayed products. For deterministic mediation, we develop a fully polynomial-time approximation scheme for a fixed number of market segments and arbitrary assortment and display capacities. For PL mediation, we develop a polynomial-time approximation scheme when both the number of market segments and display capacity are fixed.

Archive of Past YRW Programs and Abstracts

Young Researchers Workshop 2025