Hongseok Namkoong

Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre et al.

The conventional recipe for maximizing model accuracy is to (1) train multiple models with various hyperparameters and (2) pick the individual model which performs best on a held-out validation set, discarding the remainder. In this paper, we revisit the second step of this procedure in the context of fine-tuning large pre-trained models, where fine-tuned models often appear to lie in a single low error basin. We show that averaging the weights of multiple models fine-tuned with different hyperparameter configurations often improves accuracy and robustness. Unlike a conventional ensemble, we may average many models without incurring any additional inference or memory costs -- we call the results "model soups." When fine-tuning large pre-trained models such as CLIP, ALIGN, and a ViT-G pre-trained on JFT, our soup recipe provides significant improvements over the best model in a hyperparameter sweep on ImageNet. The resulting ViT-G model, which attains 90.94% top-1 accuracy on ImageNet, achieved a new state of the art. Furthermore, we show that the model soup approach extends to multiple image classification and natural language processing tasks, improves out-of-distribution performance, and improves zero-shot performance on new downstream tasks. Finally, we analytically relate the performance similarity of weight-averaging and logit-ensembling to flatness of the loss and confidence of the predictions, and validate this relation empirically. Code is available at https://github.com/mlfoundations/model-soups.

Thomas P. Zollo, Andrew Wei Tung Siah, Naimeng Ye et al.

As LLMs become capable of complex tasks, there is growing potential for personalized interactions tailored to the subtle and idiosyncratic preferences of the user. We present a public benchmark, PersonalLLM, focusing on adapting LLMs to provide maximal benefits for a particular user. Departing from existing alignment benchmarks that implicitly assume uniform preferences, we curate open-ended prompts paired with many high-quality answers over which users would be expected to display heterogeneous latent preferences. Instead of persona-prompting LLMs based on high-level attributes (e.g., user's race or response length), which yields homogeneous preferences relative to humans, we develop a method that can simulate a large user base with diverse preferences from a set of pre-trained reward models. Our dataset and generated personalities offer an innovative testbed for developing personalization algorithms that grapple with continual data sparsity--few relevant feedback from the particular user--by leveraging historical data from other (similar) users. We explore basic in-context learning and meta-learning baselines to illustrate the utility of PersonalLLM and highlight the need for future methodological development. Our dataset is available at https://huggingface.co/datasets/namkoong-lab/PersonalLLM

Hongseok Namkoong, Yuanzhe Ma, Peter W. Glynn

The performance of decision policies and prediction models often deteriorates when applied to environments different from the ones seen during training. To ensure reliable operation, we analyze the stability of a system under distribution shift, which is defined as the smallest change in the underlying environment that causes the system's performance to deteriorate beyond a permissible threshold. In contrast to standard tail risk measures and distributionally robust losses that require the specification of a plausible magnitude of distribution shift, the stability measure is defined in terms of a more intuitive quantity: the level of acceptable performance degradation. We develop a minimax optimal estimator of stability and analyze its convergence rate, which exhibits a fundamental phase shift behavior. Our characterization of the minimax convergence rate shows that evaluating stability against large performance degradation incurs a statistical cost. Empirically, we demonstrate the practical utility of our stability framework by using it to compare system designs on problems where robustness to distribution shift is critical.

Jimmy Wang, Ethan Che, Daniel R. Jiang et al.

Innovations across science and industry are evaluated using randomized trials (a.k.a. A/B tests). While simple and robust, such static designs are inefficient or infeasible for testing many hypotheses. Adaptive designs can greatly improve statistical power in theory, but they have seen limited adoption due to their fragility in practice. We present a benchmark for adaptive experimentation based on real-world datasets, highlighting prominent practical challenges to operationalizing adaptivity: non-stationarity, batched/delayed feedback, multiple outcomes and objectives, and external validity. Our benchmark aims to spur methodological development that puts practical performance (e.g., robustness) as a central concern, rather than mathematical guarantees on contrived instances. We release an open source library, AExGym, which is designed with modularity and extensibility in mind to allow experimentation practitioners to develop custom environments and algorithms.

Tiffany Tianhui Cai, Hongseok Namkoong, Steve Yadlowsky

Prediction models can perform poorly when deployed to target distributions different from the training distribution. To understand these operational failure modes, we develop a method, called DIstribution Shift DEcomposition (DISDE), to attribute a drop in performance to different types of distribution shifts. Our approach decomposes the performance drop into terms for 1) an increase in harder but frequently seen examples from training, 2) changes in the relationship between features and outcomes, and 3) poor performance on examples infrequent or unseen during training. These terms are defined by fixing a distribution on $X$ while varying the conditional distribution of $Y \mid X$ between training and target, or by fixing the conditional distribution of $Y \mid X$ while varying the distribution on $X$. In order to do this, we define a hypothetical distribution on $X$ consisting of values common in both training and target, over which it is easy to compare $Y \mid X$ and thus predictive performance. We estimate performance on this hypothetical distribution via reweighting methods. Empirically, we show how our method can 1) inform potential modeling improvements across distribution shifts for employment prediction on tabular census data, and 2) help to explain why certain domain adaptation methods fail to improve model performance for satellite image classification.

Mike Li, Daksh Mittal, Hongseok Namkoong et al.

The performance of ML models degrades when the training population is different from that seen under operation. Towards assessing distributional robustness, we study the worst-case performance of a model over all subpopulations of a given size, defined with respect to core attributes Z. This notion of robustness can consider arbitrary (continuous) attributes Z, and automatically accounts for complex intersectionality in disadvantaged groups. We develop a scalable yet principled two-stage estimation procedure that can evaluate the robustness of state-of-the-art models. We prove that our procedure enjoys several finite-sample convergence guarantees, including dimension-free convergence. Instead of overly conservative notions based on Rademacher complexities, our evaluation error depends on the dimension of Z only through the out-of-sample error in estimating the performance conditional on Z. On real datasets, we demonstrate that our method certifies the robustness of a model and prevents deployment of unreliable models.

Tianyu Wang, Jiashuo Liu, Peng Cui et al.

Different distribution shifts require different interventions, and algorithms must be grounded in the specific shifts they address. However, methodological development for robust algorithms typically relies on structural assumptions that lack empirical validation. Advocating for an empirically grounded data-driven approach to algorithm development, we build an empirical testbed comprising natural shifts across 8 tabular datasets, 172 distribution pairs over 45 methods and 90,000 method configurations encompassing empirical risk minimization and distributionally robust optimization (DRO) methods. We find $Y|X$-shifts are most prevalent in our testbed, in stark contrast to the heavy focus on $X$ (covariate)-shifts in the ML literature, and that the performance of robust algorithms is no better than that of vanilla methods. To understand why, we conduct an in-depth empirical analysis of DRO methods and find that underlooked implementation details -- such as the choice of underlying model class (e.g., LightGBM) and hyperparameter selection -- have a bigger impact on performance than the ambiguity set or its radius. We illustrate via case studies how a data-driven, inductive understanding of distribution shifts can provide a new approach to algorithm development.

Naimeng Ye, Hongseok Namkoong

Intelligent agents must be able to articulate its own uncertainty. In this work, we show that pre-trained sequence models are naturally capable of probabilistic reasoning over exchangeable data points -- forming informed beliefs and sharpening them as it gathers more information. A sequence model learns the relationship between observations, which differs from typical Bayesian models that quantify uncertainty over latent parameters through priors and likelihoods (e.g., topic models). Despite the apparent difference, we illustrate how exchangeable sequence modeling provides a valid Bayesian model by going back to De Finetti's classical predictive view of probabilistic reasoning: uncertainty comes from data that has not been observed yet, rather than latent parameters. From this perspective, pre-training autoregressive models is equivalent to formulating informed beliefs based on prior observations ("empirical Bayes"), and forward generation is equivalent to simulating instantiations of an environment ("posterior inference"). In particular, exchangeable sequence models can explicitly perform statistical inference; epistemic uncertainty over latent environments is captured by variation in predicted future observations. Formally, we show the sequence prediction loss controls the quality of uncertainty quantification, and propose several approaches for encoding exchangeability in sequence model architectures: data augmentation, regularization, and causal masking.

Ethan Che, Hongseok Namkoong

Standard bandit algorithms that assume continual reallocation of measurement effort are challenging to implement due to delayed feedback and infrastructural/organizational difficulties. Motivated by practical instances involving a handful of reallocation epochs in which outcomes are measured in batches, we develop a computation-driven adaptive experimentation framework that can flexibly handle batching. Our main observation is that normal approximations, which are universal in statistical inference, can also guide the design of adaptive algorithms. By deriving a Gaussian sequential experiment, we formulate a dynamic program that can leverage prior information on average rewards. Instead of the typical theory-driven paradigm, we leverage computational tools and empirical benchmarking for algorithm development. In particular, our empirical analysis highlights a simple yet effective algorithm, Residual Horizon Optimization, which iteratively solves a planning problem using stochastic gradient descent. Our approach significantly improves statistical power over standard methods, even when compared to Bayesian bandit algorithms (e.g., Thompson sampling) that require full distributional knowledge of individual rewards. Overall, we expand the scope of adaptive experimentation to settings that are difficult for standard methods, involving limited adaptivity, low signal-to-noise ratio, and unknown reward distributions.

Ethan Che, Jing Dong, Hongseok Namkoong

Queuing network control is essential for managing congestion in job-processing systems such as service systems, communication networks, and manufacturing processes. Despite growing interest in applying reinforcement learning (RL) techniques, queueing network control poses distinct challenges, including high stochasticity, large state and action spaces, and lack of stability. To tackle these challenges, we propose a scalable framework for policy optimization based on differentiable discrete event simulation. Our main insight is that by implementing a well-designed smoothing technique for discrete event dynamics, we can compute pathwise policy gradients for large-scale queueing networks using auto-differentiation software (e.g., Tensorflow, PyTorch) and GPU parallelization. Through extensive empirical experiments, we observe that our policy gradient estimators are several orders of magnitude more accurate than typical REINFORCE-based estimators. In addition, We propose a new policy architecture, which drastically improves stability while maintaining the flexibility of neural-network policies. In a wide variety of scheduling and admission control tasks, we demonstrate that training control policies with pathwise gradients leads to a 50-1000x improvement in sample efficiency over state-of-the-art RL methods. Unlike prior tailored approaches to queueing, our methods can flexibly handle realistic scenarios, including systems operating in non-stationary environments and those with non-exponential interarrival/service times.

Brian Hsu, Xiaotong Chen, Ying Han et al.

As AI-based decision systems proliferate, their successful operationalization requires balancing multiple desiderata: predictive performance, disparity across groups, safeguarding sensitive group attributes (e.g., race), and engineering cost. We present a holistic framework for evaluating and contextualizing fairness interventions with respect to the above desiderata. The two key points of practical consideration are \emph{where} (pre-, in-, post-processing) and \emph{how} (in what way the sensitive group data is used) the intervention is introduced. We demonstrate our framework with a case study on predictive parity. In it, we first propose a novel method for achieving predictive parity fairness without using group data at inference time via distibutionally robust optimization. Then, we showcase the effectiveness of these methods in a benchmarking study of close to 400 variations across two major model types (XGBoost vs. Neural Net), ten datasets, and over twenty unique methodologies. Methodological insights derived from our empirical study inform the practical design of ML workflow with fairness as a central concern. We find predictive parity is difficult to achieve without using group data, and despite requiring group data during model training (but not inference), distributionally robust methods we develop provide significant Pareto improvement. Moreover, a plain XGBoost model often Pareto-dominates neural networks with fairness interventions, highlighting the importance of model inductive bias.

Ethan Che, Daniel R. Jiang, Hongseok Namkoong et al.

Real-world experiments involve batched & delayed feedback, non-stationarity, multiple objectives & constraints, and (often some) personalization. Tailoring adaptive methods to address these challenges on a per-problem basis is infeasible, and static designs remain the de facto standard. Focusing on short-horizon ($\le 10$) adaptive experiments, we move away from bespoke algorithms and present a mathematical programming formulation that can flexibly incorporate a wide range of objectives, constraints, and statistical procedures. We formulating a dynamic program based on central limit approximations, which enables the use of scalable optimization methods based on auto-differentiation and GPU parallelization. To evaluate our framework, we implement a simple heuristic planning method ("solver") and benchmark it across hundreds of problem instances involving non-stationarity, personalization, and multiple objectives & constraints. Unlike bespoke methods (e.g., Thompson sampling variants), our mathematical programming framework provides consistent gains over static randomized control trials and exhibits robust performance across problem instances.

Tommaso Castellani, Naimeng Ye, Daksh Mittal et al.

AI agents increasingly rely on external tools to solve complex, long-horizon tasks. Advancing such agents requires reproducible evaluation and large-scale training in controllable, diverse, and realistic tool-use environments. However, real-world APIs are limited in availability, domain coverage, and stability, often requiring access keys and imposing rate limits, which render them impractical for stable evaluation or scalable training. To address these challenges, we introduce SynthTools, a flexible and scalable framework for generating synthetic tool ecosystems. Our framework consists of three core components: Tool Generation for automatic and scalable creation of diverse tools, Tool Simulation to emulate realistic tool behaviors, and Tool Audit to ensure correctness and consistency of tool simulation. To illustrate its scalability, we show that SynthTools can readily produce toolsets that span twice as many domains and twice as many tools per domain as prior work. Furthermore, the tool simulation and tool audit components demonstrate strong reliability, achieving $94\%$ and $99\%$ accuracy respectively. Finally, we construct downstream tasks from the generated tools that even state-of-the-art models struggle to complete. By enabling scalable, diverse, and reliable tool ecosystems, SynthTools provides a practical path toward large-scale training and stable evaluation of tool-use agents. Our code is available at https://github.com/namkoong-lab/SynthTools.

Hongseok Namkoong

Empirical likelihood is an attractive inferential framework that respects natural parameter boundaries, but existing approaches typically require smoothness of the functional and miscalibrate substantially when these assumptions are violated. For the optimal-value functional central to policy evaluation, smoothness holds only when the optimum is unique -- a condition that fails exactly when rigorous inference is most needed where more complex policies have modest gains. In this work, we develop a bootstrap empirical likelihood method for partially nonsmooth functionals. Our analytic workhorse is a geometric reduction of the profile likelihood to the distance between the score mean and a level set whose shape (a tangent cone given by nonsmoothness patterns) determines the asymptotic distribution. Unlike the classical proof technology based on Taylor expansions on the dual optima, our geometric approach leverages properties of a deterministic convex program and can directly apply to nonsmooth functionals. Since the ordinary bootstrap is not valid in the presence of nonsmoothness, we derive a corrected multiplier bootstrap approach that adapts to the unknown level-set geometry.

Thomson Yen, Julian Poeltl, Harshith Srinivas Gear et al.

LLM agents are increasingly expected to carry out end-to-end workflows, producing complete artifacts from high-level user instructions. To meet enterprise needs, frontier AI labs have developed agents that can construct entire spreadsheets from scratch. This is especially relevant in finance, where core workflows such as financial modeling, forecasting, and scenario analysis are commonly conducted through spreadsheets. Yet, existing spreadsheet benchmarks do not measure this advanced capability, focusing instead on question-answering or single-formula edits. To address this gap, we provide one of the first evaluations of agents on end-to-end spreadsheet tasks, focusing on economically critical financial workflows such as modeling and scenario analysis. Since deliverables therein are routinely reviewed and revised by multiple stakeholders, judging their quality necessarily involves high-level criteria such as readability or ease of modification. To reflect the multidimensional nature of solution quality, we develop an evaluation taxonomy comprising three dimensions: Accuracy, Formula, and Format, each comprising fine-grained criteria that reflect professional standards. The Claude family leads the benchmark and produces the most professional-looking outputs in our qualitative review, but even the strongest agents frequently fall short of professional finance standards and degrade sharply as the difficulty increases beyond a few chained calculations. This suggests that current agents are not yet able to reliably produce professional-quality spreadsheets at the level of complexity real-world workflows demand.

Ang Li, Haozhe Chen, Hongseok Namkoong et al.

The use of large language models (LLMs) to simulate human behavior has gained significant attention, particularly through personas that approximate individual characteristics. Persona-based simulations hold promise for transforming disciplines that rely on population-level feedback, including social science, economic analysis, marketing research, and business operations. Traditional methods to collect realistic persona data face significant challenges. They are prohibitively expensive and logistically challenging due to privacy constraints, and often fail to capture multi-dimensional attributes, particularly subjective qualities. Consequently, synthetic persona generation with LLMs offers a scalable, cost-effective alternative. However, current approaches rely on ad hoc and heuristic generation techniques that do not guarantee methodological rigor or simulation precision, resulting in systematic biases in downstream tasks. Through extensive large-scale experiments including presidential election forecasts and general opinion surveys of the U.S. population, we reveal that these biases can lead to significant deviations from real-world outcomes. Our findings underscore the need to develop a rigorous science of persona generation and outline the methodological innovations, organizational and institutional support, and empirical foundations required to enhance the reliability and scalability of LLM-driven persona simulations. To support further research and development in this area, we have open-sourced approximately one million generated personas, available for public access and analysis at https://huggingface.co/datasets/Tianyi-Lab/Personas.

Jiashuo Liu, Jiayun Wu, Chunjie Wu et al.

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation ($N=100,000$ iterations) is used to approximate the statistically robust Expected Win Score ($E[S_m]$), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity ($T_k$), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

Liang Hu, Jianpeng Jiao, Jiashuo Liu et al.

Search has emerged as core infrastructure for LLM-based agents and is widely viewed as critical on the path toward more general intelligence. Finance is a particularly demanding proving ground: analysts routinely conduct complex, multi-step searches over time-sensitive, domain-specific data, making it ideal for assessing both search proficiency and knowledge-grounded reasoning. Yet no existing open financial datasets evaluate data searching capability of end-to-end agents, largely because constructing realistic, complicated tasks requires deep financial expertise and time-sensitive data is hard to evaluate. We present FinSearchComp, the first fully open-source agent benchmark for realistic, open-domain financial search and reasoning. FinSearchComp comprises three tasks -- Time-Sensitive Data Fetching, Simple Historical Lookup, and Complex Historical Investigation -- closely reproduce real-world financial analyst workflows. To ensure difficulty and reliability, we engage 70 professional financial experts for annotation and implement a rigorous multi-stage quality-assurance pipeline. The benchmark includes 635 questions spanning global and Greater China markets, and we evaluate 21 models (products) on it. Grok 4 (web) tops the global subset, approaching expert-level accuracy. DouBao (web) leads on the Greater China subset. Experimental analyses show that equipping agents with web search and financial plugins substantially improves results on FinSearchComp, and the country origin of models and tools impact performance significantly.By aligning with realistic analyst tasks and providing end-to-end evaluation, FinSearchComp offers a professional, high-difficulty testbed for complex financial search and reasoning.

Jiashuo Liu, Tianyu Wang, Henry Lam et al.

We introduce dro, an open-source Python library for distributionally robust optimization (DRO) for regression and classification problems. The library implements 14 DRO formulations and 9 backbone models, enabling 79 distinct DRO methods. Furthermore, dro is compatible with both scikit-learn and PyTorch. Through vectorization and optimization approximation techniques, dro reduces runtime by 10x to over 1000x compared to baseline implementations on large-scale datasets. Comprehensive documentation is available at https://python-dro.org.

Matthew O'Kelly, Aman Sinha, Justin Norden et al.

Modern treatments for Type 1 diabetes (T1D) use devices known as artificial pancreata (APs), which combine an insulin pump with a continuous glucose monitor (CGM) operating in a closed-loop manner to control blood glucose levels. In practice, poor performance of APs (frequent hyper- or hypoglycemic events) is common enough at a population level that many T1D patients modify the algorithms on existing AP systems with unregulated open-source software. Anecdotally, the patients in this group have shown superior outcomes compared with standard of care, yet we do not understand how safe any AP system is since adverse outcomes are rare. In this paper, we construct generative models of individual patients' physiological characteristics and eating behaviors. We then couple these models with a T1D simulator approved for pre-clinical trials by the FDA. Given the ability to simulate patient outcomes in-silico, we utilize techniques from rare-event simulation theory in order to efficiently quantify the performance of a device with respect to a particular patient. We show a 72,000$\times$ speedup in simulation speed over real-time and up to 2-10 times increase in the frequency which we are able to sample adverse conditions relative to standard Monte Carlo sampling. In practice our toolchain enables estimates of the likelihood of hypoglycemic events with approximately an order of magnitude fewer simulations.

Jimmy Wang, Thomas Zollo, Richard Zemel et al.

Eliciting information to reduce uncertainty about a latent entity is a critical task in many application domains, e.g., assessing individual student learning outcomes, diagnosing underlying diseases, or learning user preferences. Though natural language is a powerful medium for this purpose, large language models (LLMs) and existing fine-tuning algorithms lack mechanisms for strategically gathering information to refine their own understanding of the latent entity. To harness the generalization power and world knowledge of LLMs in developing effective information-gathering strategies, we propose an adaptive elicitation framework that actively reduces uncertainty on the latent entity. Since probabilistic modeling of an abstract latent entity is difficult, our framework adopts a predictive view of uncertainty, using a meta-learned language model to simulate future observations and enable scalable uncertainty quantification over complex natural language. Through autoregressive forward simulation, our model quantifies how new questions reduce epistemic uncertainty, enabling the development of sophisticated information-gathering strategies to choose the most informative next queries. In experiments on the 20 questions game, dynamic opinion polling, and adaptive student assessment, our method consistently outperforms baselines in identifying critical unknowns and improving downstream predictions, illustrating the promise of strategic information gathering in natural language settings.

Kelly W. Zhang, Tiffany Tianhui Cai, Hongseok Namkoong et al.

We introduce a framework for Thompson sampling (TS) contextual bandit algorithms, in which the algorithm's ability to quantify uncertainty and make decisions depends on the quality of a generative model that is learned offline. Instead of viewing uncertainty in the environment as arising from unobservable latent parameters, our algorithm treats uncertainty as stemming from missing, but potentially observable outcomes (including both future and counterfactual outcomes). If these outcomes were all observed, one could simply make decisions using an "oracle" policy fit on the complete dataset. Inspired by this conceptualization, at each decision-time, our algorithm uses a generative model to probabilistically impute missing outcomes, fits a policy using the imputed complete dataset, and uses that policy to select the next action. We formally show that this algorithm is a generative formulation of TS and establish a state-of-the-art regret bound. Notably, our regret bound depends on the generative model only through the quality of its offline prediction loss, and applies to any method of fitting the "oracle" policy.

Thomson Yen, Andrew Wei Tung Siah, Haozhe Chen et al.

Careful curation of data sources can significantly improve the performance of LLM pre-training, but predominant approaches rely heavily on intuition or costly trial-and-error, making them difficult to generalize across different data domains and downstream tasks. Although scaling laws can provide a principled and general approach for data curation, standard deterministic extrapolation from small-scale experiments to larger scales requires strong assumptions on the reliability of such extrapolation, whose brittleness has been highlighted in prior works. In this paper, we introduce a $\textit{probabilistic extrapolation framework}$ for data mixture optimization that avoids rigid assumptions and explicitly models the uncertainty in performance across decision variables. We formulate data curation as a sequential decision-making problem$\unicode{x2013}$multi-fidelity, multi-scale Bayesian optimization$\unicode{x2013}$where $\{$data mixtures, model scale, training steps$\}$ are adaptively selected to balance training cost and potential information gain. Our framework naturally gives rise to algorithm prototypes that leverage noisy information from inexpensive experiments to systematically inform costly training decisions. To accelerate methodological progress, we build a simulator based on 472 language model pre-training runs with varying data compositions from the SlimPajama dataset. We observe that even simple kernels and acquisition functions can enable principled decisions across training models from 20M to 1B parameters and achieve $\textbf{2.6x}$ and $\textbf{3.3x}$ speedups compared to multi-fidelity BO and random search baselines. Taken together, our framework underscores potential efficiency gains achievable by developing principled and transferable data mixture optimization methods.

Daksh Mittal, Ang Li, Tzu-Ching Yen et al.

Autoregressive models have emerged as a powerful framework for modeling exchangeable sequences - i.i.d. observations when conditioned on some latent factor - enabling direct modeling of uncertainty from missing data (rather than a latent). Motivated by the critical role posterior inference plays as a subroutine in decision-making (e.g., active learning, bandits), we study the inferential and architectural inductive biases that are most effective for exchangeable sequence modeling. For the inference stage, we highlight a fundamental limitation of the prevalent single-step generation approach: inability to distinguish between epistemic and aleatoric uncertainty. Instead, a long line of works in Bayesian statistics advocates for multi-step autoregressive generation; we demonstrate this "correct approach" enables superior uncertainty quantification that translates into better performance on downstream decision-making tasks. This naturally leads to the next question: which architectures are best suited for multi-step inference? We identify a subtle yet important gap between recently proposed Transformer architectures for exchangeable sequences (Muller et al., 2022; Nguyen & Grover, 2022; Ye & Namkoong, 2024), and prove that they in fact cannot guarantee exchangeability despite introducing significant computational overhead. We illustrate our findings using controlled synthetic settings, demonstrating how custom architectures can significantly underperform standard causal masks, underscoring the need for new architectural innovations.

Tiffany Tianhui Cai, Yuri Fonseca, Kaiwen Hou et al.

Popular debiased estimation methods for causal inference -- such as augmented inverse propensity weighting and targeted maximum likelihood estimation -- enjoy desirable asymptotic properties like statistical efficiency and double robustness but they can produce unstable estimates when there is limited overlap between treatment and control, requiring additional assumptions or ad hoc adjustments in practice (e.g., truncating propensity scores). In contrast, simple plug-in estimators are stable but lack desirable asymptotic properties. We propose a novel debiasing approach that achieves the best of both worlds, producing stable plug-in estimates with desirable asymptotic properties. Our constrained learning framework solves for the best plug-in estimator under the constraint that the first-order error with respect to the plugged-in quantity is zero, and can leverage flexible model classes including neural networks and tree ensembles. In several experimental settings, including ones in which we handle text-based covariates by fine-tuning language models, our constrained learning-based estimator outperforms basic versions of one-step estimation and targeting in challenging settings with limited overlap between treatment and control, and performs similarly otherwise.

Gilbert Yang, Yaqin Chen, Thomson Yen et al.

To reliably navigate ever-shifting real-world environments, agents must grapple with incomplete knowledge and adapt their behavior through experience. However, current evaluations largely focus on tasks that leave no ambiguity, and do not measure agents' ability to adaptively learn and reason through the experiences they accrued. We exemplify the need for this in-context experiential learning in a product recommendation context, where agents must navigate shifting customer preferences and product landscapes through natural language dialogue. We curate a benchmark for experiential learning and active exploration (BELA) that combines (1) rich real-world products from Amazon, (2) a diverse collection of user personas to represent heterogeneous yet latent preferences, and (3) a LLM user simulator powered by the persona to create rich interactive trajectories. We observe that current frontier models struggle to meaningfully improve across episodes, underscoring the need for agentic systems with strong in-context learning capabilities.

Yuanzhe Ma, Hongseok Namkoong

Limited overlap between treated and control groups is a key challenge in observational analysis. Standard approaches like trimming importance weights can reduce variance but introduce a fundamental bias. We propose a sensitivity framework for contextualizing findings under limited overlap, where we assess how irregular the outcome function has to be in order for the main finding to be invalidated. Our approach is based on worst-case confidence bounds on the bias introduced by standard trimming practices, under explicit assumptions necessary to extrapolate counterfactual estimates from regions of overlap to those without. Empirically, we demonstrate how our sensitivity framework protects against spurious findings by quantifying uncertainty in regions with limited overlap.

Yuta Kobayashi, Zilin Jing, Jiayu Yao et al.

Active feature acquisition (AFA) is a sequential decision-making problem where the goal is to improve model performance for test instances by adaptively selecting which features to acquire. In practice, AFA methods often learn from retrospective data with systematic missingness in the features and limited task-specific labels. Most prior work addresses acquisition for a single predetermined task, limiting scalability. To address this limitation, we formalize the meta-AFA problem, where the goal is to learn acquisition policies across various tasks. We introduce Learning-to-Measure (L2M), which consists of i) reliable uncertainty quantification over unseen tasks, and ii) an uncertainty-guided greedy feature acquisition agent that maximizes conditional mutual information. We demonstrate a sequence-modeling or autoregressive pre-training approach that underpins reliable uncertainty quantification for tasks with arbitrary missingness. L2M operates directly on datasets with retrospective missingness and performs the meta-AFA task in-context, eliminating per-task retraining. Across synthetic and real-world tabular benchmarks, L2M matches or surpasses task-specific baselines, particularly under scarce labels and high missingness.

Yanlin Qu, Hongseok Namkoong, Assaf Zeevi

Thompson Sampling is one of the most widely used and studied bandit algorithms, known for its simple structure, low regret performance, and solid theoretical guarantees. Yet, in stark contrast to most other families of bandit algorithms, the exact mechanism through which posterior sampling (as introduced by Thompson) is able to "properly" balance exploration and exploitation, remains a mystery. In this paper we show that the core insight to address this question stems from recasting Thompson Sampling as an online optimization algorithm. To distill this, a key conceptual tool is introduced, which we refer to as "faithful" stationarization of the regret formulation. Essentially, the finite horizon dynamic optimization problem is converted into a stationary counterpart which "closely resembles" the original objective (in contrast, the classical infinite horizon discounted formulation, that leads to the Gittins index, alters the problem and objective in too significant a manner). The newly crafted time invariant objective can be studied using Bellman's principle which leads to a time invariant optimal policy. When viewed through this lens, Thompson Sampling admits a simple online optimization form that mimics the structure of the Bellman-optimal policy, and where greediness is regularized by a measure of residual uncertainty based on point-biserial correlation. This answers the question of how Thompson Sampling balances exploration-exploitation, and moreover, provides a principled framework to study and further improve Thompson's original idea.

Daksh Mittal, Shunri Zheng, Jing Dong et al.

While queueing network models are powerful tools for analyzing service systems, they traditionally require substantial human effort and domain expertise to construct. To make this modeling approach more scalable and accessible, we propose a data-driven framework for queueing network modeling and simulation based on autoregressive sequence models trained on event-stream data. Instead of explicitly specifying arrival processes, service mechanisms, or routing logic, our approach learns the conditional distributions of event types and event times, recasting the modeling task as a problem of sequence distribution learning. We show that Transformer-style architectures can effectively parameterize these distributions, enabling automated construction of high-fidelity simulators. As a proof of concept, we validate our framework on event tables generated from diverse queueing networks, showcasing its utility in simulation, uncertainty quantification, and counterfactual evaluation. Leveraging advances in artificial intelligence and the growing availability of data, our framework takes a step toward more automated, data-driven modeling pipelines to support broader adoption of queueing network models across service domains.

Daksh Mittal, Yuanzhe Ma, Shalmali Joshi et al.

Ground truth labels/outcomes are critical for advancing scientific and engineering applications, e.g., evaluating the treatment effect of an intervention or performance of a predictive model. Since randomly sampling inputs for labeling can be prohibitively expensive, we introduce an adaptive labeling framework where measurement effort can be reallocated in batches. We formulate this problem as a Markov decision process where posterior beliefs evolve over time as batches of labels are collected (state transition), and batches (actions) are chosen to minimize uncertainty at the end of data collection. We design a computational framework that is agnostic to different uncertainty quantification approaches including those based on deep learning, and allows a diverse array of policy gradient approaches by relying on continuous policy parameterizations. On real and synthetic datasets, we demonstrate even a one-step lookahead policy can substantially outperform common adaptive labeling heuristics, highlighting the virtue of planning. On the methodological side, we note that standard REINFORCE-style policy gradient estimators can suffer high variance since they rely only on zeroth order information. We propose a direct backpropagation-based approach, Smoothed-Autodiff, based on a carefully smoothed version of the original non-differentiable MDP. Our method enjoys low variance at the price of introducing bias, and we theoretically and empirically show that this trade-off can be favorable.

Brian Hsu, Cyrus DiCiccio, Natesh Sivasubramoniapillai et al.

Fairness research in machine learning often centers on ensuring equitable performance of individual models. However, real-world recommendation systems are built on multiple models and even multiple stages, from candidate retrieval to scoring and serving, which raises challenges for responsible development and deployment. This system-level view, as highlighted by regulations like the EU AI Act, necessitates moving beyond auditing individual models as independent entities. We propose a holistic framework for modeling system-level fairness, focusing on the end-utility delivered to diverse user groups, and consider interactions between components such as retrieval and scoring models. We provide formal insights on the limitations of focusing solely on model-level fairness and highlight the need for alternative tools that account for heterogeneity in user preferences. To mitigate system-level disparities, we adapt closed-box optimization tools (e.g., BayesOpt) to jointly optimize utility and equity. We empirically demonstrate the effectiveness of our proposed framework on synthetic and real datasets, underscoring the need for a system-level framework.

Jiung Lee, Hongseok Namkoong, Yibo Zeng

To leverage prediction models to make optimal scheduling decisions in service systems, we must understand how predictive errors impact congestion due to externalities on the delay of other jobs. Motivated by applications where prediction models interact with human servers (e.g., content moderation), we consider a large queueing system comprising of many single server queues where the class of a job is estimated using a prediction model. By characterizing the impact of mispredictions on congestion cost in heavy traffic, we design an index-based policy that incorporates the predicted class information in a near-optimal manner. Our theoretical results guide the design of predictive models by providing a simple model selection procedure with downstream queueing performance as a central concern, and offer novel insights on how to design queueing systems with AI-based triage. We illustrate our framework on a content moderation task based on real online comments, where we construct toxicity classifiers by finetuning large language models.

Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim et al.

Large pre-trained models such as CLIP or ALIGN offer consistent accuracy across a range of data distributions when performing zero-shot inference (i.e., without fine-tuning on a specific dataset). Although existing fine-tuning methods substantially improve accuracy on a given target distribution, they often reduce robustness to distribution shifts. We address this tension by introducing a simple and effective method for improving robustness while fine-tuning: ensembling the weights of the zero-shot and fine-tuned models (WiSE-FT). Compared to standard fine-tuning, WiSE-FT provides large accuracy improvements under distribution shift, while preserving high accuracy on the target distribution. On ImageNet and five derived distribution shifts, WiSE-FT improves accuracy under distribution shift by 4 to 6 percentage points (pp) over prior work while increasing ImageNet accuracy by 1.6 pp. WiSE-FT achieves similarly large robustness gains (2 to 23 pp) on a diverse set of six further distribution shifts, and accuracy gains of 0.8 to 3.3 pp compared to standard fine-tuning on seven commonly used transfer learning datasets. These improvements come at no additional computational cost during fine-tuning or inference.

Hongseok Namkoong, Samuel Daulton, Eytan Bakshy

Thompson sampling (TS) has emerged as a robust technique for contextual bandit problems. However, TS requires posterior inference and optimization for action generation, prohibiting its use in many online platforms where latency and ease of deployment are of concern. We operationalize TS by proposing a novel imitation-learning-based algorithm that distills a TS policy into an explicit policy representation, allowing fast decision-making and easy deployment in mobile and server-based environments. Using batched data collected under the imitation policy, our algorithm iteratively performs offline updates to the TS policy, and learns a new explicit policy representation to imitate it. Empirically, our imitation policy achieves performance comparable to batch TS while allowing more than an order of magnitude reduction in decision-time latency. Buoyed by low latency and simplicity of implementation, our algorithm has been successfully deployed in multiple video upload systems for Meta. Using a randomized controlled trial, we show our algorithm resulted in significant improvements in video quality and watch time.

John Duchi, Tatsunori Hashimoto, Hongseok Namkoong

While modern large-scale datasets often consist of heterogeneous subpopulations -- for example, multiple demographic groups or multiple text corpora -- the standard practice of minimizing average loss fails to guarantee uniformly low losses across all subpopulations. We propose a convex procedure that controls the worst-case performance over all subpopulations of a given size. Our procedure comes with finite-sample (nonparametric) convergence guarantees on the worst-off subpopulation. Empirically, we observe on lexical similarity, wine quality, and recidivism prediction tasks that our worst-case procedure learns models that do well against unseen subpopulations.

Sookyo Jeong, Hongseok Namkoong

Study populations are typically sampled from limited points in space and time, and marginalized groups are underrepresented. To assess the external validity of randomized and observational studies, we propose and evaluate the worst-case treatment effect (WTE) across all subpopulations of a given size, which guarantees positive findings remain valid over subpopulations. We develop a semiparametrically efficient estimator for the WTE that analyzes the external validity of the augmented inverse propensity weighted estimator for the average treatment effect. Our cross-fitting procedure leverages flexible nonparametric and machine learning-based estimates of nuisance parameters and is a regular root-$n$ estimator even when nuisance estimates converge more slowly. On real examples where external validity is of core concern, our proposed framework guards against brittle findings that are invalidated by unanticipated population shifts.

Hongseok Namkoong, Ramtin Keramati, Steve Yadlowsky et al.

When observed decisions depend only on observed features, off-policy policy evaluation (OPE) methods for sequential decision making problems can estimate the performance of evaluation policies before deploying them. This assumption is frequently violated due to unobserved confounders, unrecorded variables that impact both the decisions and their outcomes. We assess robustness of OPE methods under unobserved confounding by developing worst-case bounds on the performance of an evaluation policy. When unobserved confounders can affect every decision in an episode, we demonstrate that even small amounts of per-decision confounding can heavily bias OPE methods. Fortunately, in a number of important settings found in healthcare, policy-making, operations, and technology, unobserved confounders may primarily affect only one of the many decisions made. Under this less pessimistic model of one-decision confounding, we propose an efficient loss-minimization-based procedure for computing worst-case bounds, and prove its statistical consistency. On two simulated healthcare examples---management of sepsis patients and developmental interventions for autistic children---where this is a reasonable model of confounding, we demonstrate that our method invalidates non-robust results and provides meaningful certificates of robustness, allowing reliable selection of policies even under unobserved confounding.

Matthew O'Kelly, Aman Sinha, Hongseok Namkoong et al.

While recent developments in autonomous vehicle (AV) technology highlight substantial progress, we lack tools for rigorous and scalable testing. Real-world testing, the $\textit{de facto}$ evaluation environment, places the public in danger, and, due to the rare nature of accidents, will require billions of miles in order to statistically validate performance claims. We implement a simulation framework that can test an entire modern autonomous driving system, including, in particular, systems that employ deep-learning perception and control algorithms. Using adaptive importance-sampling methods to accelerate rare-event probability evaluation, we estimate the probability of an accident under a base distribution governing standard traffic behavior. We demonstrate our framework on a highway scenario, accelerating system evaluation by $2$-$20$ times over naive Monte Carlo sampling methods and $10$-$300 \mathsf{P}$ times (where $\mathsf{P}$ is the number of processors) over real-world testing.

John Duchi, Hongseok Namkoong

A common goal in statistics and machine learning is to learn models that can perform well against distributional shifts, such as latent heterogeneous subpopulations, unknown covariate shifts, or unmodeled temporal effects. We develop and analyze a distributionally robust stochastic optimization (DRO) framework that learns a model providing good performance against perturbations to the data-generating distribution. We give a convex formulation for the problem, providing several convergence guarantees. We prove finite-sample minimax upper and lower bounds, showing that distributional robustness sometimes comes at a cost in convergence rates. We give limit theorems for the learned parameters, where we fully specify the limiting distribution so that confidence intervals can be computed. On real tasks including generalizing to unknown subpopulations, fine-grained recognition, and providing good tail performance, the distributionally robust approach often exhibits improved performance.

Tatsunori B. Hashimoto, Megha Srivastava, Hongseok Namkoong et al.

Machine learning models (e.g., speech recognizers) are usually trained to minimize average loss, which results in representation disparity---minority groups (e.g., non-native speakers) contribute less to the training objective and thus tend to suffer higher loss. Worse, as model accuracy affects user retention, a minority group can shrink over time. In this paper, we first show that the status quo of empirical risk minimization (ERM) amplifies representation disparity over time, which can even make initially fair models unfair. To mitigate this, we develop an approach based on distributionally robust optimization (DRO), which minimizes the worst case risk over all distributions close to the empirical distribution. We prove that this approach controls the risk of the minority group at each time step, in the spirit of Rawlsian distributive justice, while remaining oblivious to the identity of the groups. We demonstrate that DRO prevents disparity amplification on examples where ERM fails, and show improvements in minority group user satisfaction in a real-world text autocomplete task.

Riccardo Volpi, Hongseok Namkoong, Ozan Sener et al.

We are concerned with learning models that generalize well to different \emph{unseen} domains. We consider a worst-case formulation over data distributions that are near the source domain in the feature space. Only using training data from a single source distribution, we propose an iterative procedure that augments the dataset with examples from a fictitious target domain that is "hard" under the current model. We show that our iterative scheme is an adaptive data augmentation method where we append adversarial examples at each iteration. For softmax losses, we show that our method is a data-dependent regularization scheme that behaves differently from classical regularizers that regularize towards zero (e.g., ridge or lasso). On digit recognition and semantic segmentation tasks, our method learns models improve performance across a range of a priori unknown target domains.

Aman Sinha, Hongseok Namkoong, Riccardo Volpi et al.

Neural networks are vulnerable to adversarial examples and researchers have proposed many heuristic attack and defense mechanisms. We address this problem through the principled lens of distributionally robust optimization, which guarantees performance under adversarial input perturbations. By considering a Lagrangian penalty formulation of perturbing the underlying data distribution in a Wasserstein ball, we provide a training procedure that augments model parameter updates with worst-case perturbations of training data. For smooth losses, our procedure provably achieves moderate levels of robustness with little computational or statistical cost relative to empirical risk minimization. Furthermore, our statistical guarantees allow us to efficiently certify robustness for the population loss. For imperceptible perturbations, our method matches or outperforms heuristic approaches.

John Duchi, Peter Glynn, Hongseok Namkoong

We study statistical inference and distributionally robust solution methods for stochastic optimization problems, focusing on confidence intervals for optimal values and solutions that achieve exact coverage asymptotically. We develop a generalized empirical likelihood framework---based on distributional uncertainty sets constructed from nonparametric $f$-divergence balls---for Hadamard differentiable functionals, and in particular, stochastic optimization problems. As consequences of this theory, we provide a principled method for choosing the size of distributional uncertainty regions to provide one- and two-sided confidence intervals that achieve exact coverage. We also give an asymptotic expansion for our distributionally robust formulation, showing how robustification regularizes problems by their variance. Finally, we show that optimizers of the distributionally robust formulations we study enjoy (essentially) the same consistency properties as those in classical sample average approximations. Our general approach applies to quickly mixing stationary sequences, including geometrically ergodic Harris recurrent Markov chains.

John Duchi, Hongseok Namkoong

We develop an approach to risk minimization and stochastic optimization that provides a convex surrogate for variance, allowing near-optimal and computationally efficient trading between approximation and estimation error. Our approach builds off of techniques for distributionally robust optimization and Owen's empirical likelihood, and we provide a number of finite-sample and asymptotic results characterizing the theoretical performance of the estimator. In particular, we show that our procedure comes with certificates of optimality, achieving (in some scenarios) faster rates of convergence than empirical risk minimization by virtue of automatically balancing bias and variance. We give corroborating empirical evidence showing that in practice, the estimator indeed trades between variance and absolute performance on a training sample, improving out-of-sample (test) performance over standard empirical risk minimization for a number of classification problems.