AISTATS 2025 论文列表

Logistic regression models are widely used in the social and behavioral sciences and in high-stakes domains, due to their simplicity and interpretability properties. At the same time, such domains are permeated by distribution shifts, where the distribution generating the data changes between training and deployment. In this paper, we study a distributionally robust logistic regression problem that seeks the model that will perform best against adversarial realizations of the data distribution drawn from a suitably constructed Wasserstein ambiguity set. Our model and solution approach differ from prior work in that we can capture settings where the likelihood of distribution shifts can vary across features, significantly broadening the applicability of our model relative to the state-of-the-art. We propose a graph-based solution approach that can be integrated into off-the-shelf optimization solvers. We evaluate the performance of our model and algorithms on numerous publicly available datasets. Our solution achieves a 408x speed-up relative to the state-of-the-art. Additionally, compared to the state-of-the-art, our model reduces average calibration error by up to 36.19% and worst-case calibration error by up to 41.70%, while increasing the average area under the ROC curve (AUC) by up to 18.02% and worst-case AUC by up to 48.37%.

Foundation models are now a major focus of leading technology organizations due to their ability to generalize across diverse tasks. Existing approaches for adapting foundation models to new applications often rely on Federated Learning (FL) and disclose the foundation model weights to clients when using it to initialize the global model. While these methods ensure client data privacy, they compromise model and information security. In this paper, we introduce Federated Learning Aggregation Biased by a Foundation Model (FedBaF), a novel method for dynamically integrating pre-trained foundation model weights during the FL aggregation phase. Unlike conventional methods, FedBaF preserves the confidentiality of the foundation model while still leveraging its power to train more accurate models, especially in non-IID and adversarial scenarios. Our comprehensive experiments use Pre-ResNet and foundation models like Vision Transformer to demonstrate that FedBaF not only matches, but often surpasses the test accuracy of traditional weight initialization methods by up to 11.4\% in IID and up to 15.8\% in non-IID settings. Additionally, FedBaF applied to a Transformer-based language model significantly reduced perplexity by up to 39.2\%.

This paper revisits the online learning approach to inverse linear optimization studied by Bärmann et al. (2017), where the goal is to infer an unknown linear objective function of an agent from sequential observations of the agent's input-output pairs. First, we provide a simple understanding of the online learning approach through its connection to online convex optimization of *Fenchel–Young losses*. As a byproduct, we present an offline guarantee on the *suboptimality loss*, which measures how well predicted objective vectors explain the agent's choices, without assuming the optimality of the agent's choices. Second, assuming that there is a gap between optimal and suboptimal objective values in the agent's decision problems, we obtain an upper bound independent of the time horizon $T$ on the sum of suboptimality and *estimate losses*, where the latter measures the quality of solutions recommended by predicted objective vectors. Interestingly, our gap-dependent analysis achieves a faster rate than the standard $O(\sqrt{T})$ regret bound by exploiting structures specific to inverse linear optimization, even though neither the loss functions nor their domains possess desirable properties, such as strong convexity.

We investigate the non-stationary stochastic linear bandit problem where the reward distribution evolves each round. Existing algorithms characterize the non-stationarity by the total variation budget $B_K$, which is the summation of the change of the consecutive feature vectors of the linear bandits over $K$ rounds. However, such a quantity only measures the non-stationarity with respect to the expectation of the reward distribution, which makes existing algorithms sub-optimal under the general non-stationary distribution setting. In this work, we propose algorithms that utilize the variance of the reward distribution as well as the $B_K$, and show that they can achieve tighter regret upper bounds. Specifically, we introduce two novel algorithms: Restarted Weighted$\text{OFUL}^+$ and Restarted $\text{SAVE}^+$. These algorithms address cases where the variance information of the rewards is known and unknown, respectively. Notably, when the total variance $V_K$ is much smaller than $K$, our algorithms outperform previous state-of-the-art results on non-stationary stochastic linear bandits under different settings. Experimental evaluations further validate the superior performance of our proposed algorithms over existing works.

This paper studies the cooperative stochastic $k$-armed bandit problem, where $m$ agents collaborate to identify the optimal action. Rather than adapting a specific single-agent algorithm, we propose a general-purpose black-box reduction that extends any single-agent algorithm to the multi-agent setting. Under mild assumptions, we prove that our black-box approach preserves the regret guarantees of the chosen algorithm, and is capable of achieving minimax-optimality up to an additive graph-dependent term. Our method applies to various bandit settings, including heavy-tailed and duelling bandits, and those with local differential privacy. Empirically, it is competitive with or outperforms specialized multi-agent algorithms.

We consider distributed kernel bandits where $N$ agents aim to collaboratively maximize an unknown reward function that lies in a reproducing kernel Hilbert space. Each agent sequentially queries the function to obtain noisy observations at the query points. Agents can share information through a central server, with the objective of minimizing regret that is accumulating over time $T$ and aggregating over agents. We develop the first algorithm that achieves the optimal regret order (as defined by centralized learning) with a communication cost that is sublinear in both $N$ and $T$. The key features of the proposed algorithm are the uniform exploration at the local agents and shared randomness with the central server. Working together with the sparse approximation of the GP model, these two key components make it possible to preserve the learning rate of the centralized setting at a diminishing rate of communication.

Large language models (LLMs) contain substantial factual knowledge which is commonly elicited by multiple-choice question-answering prompts. Internally, such models process the prompt through multiple transformer layers, building varying representations of the problem within its hidden states. Ultimately, however, only the hidden state corresponding to the final layer and token position is used to predict the answer label. In this work, we propose instead to learn a small separate neural network predictor module on a collection of training questions, that take the hidden states from all the layers at the last temporal position as input and outputs predictions. In effect, such a framework disentangles the representational abilities of LLMs from their predictive abilities. On a collection of hard benchmarks, our method achieves considerable improvements in performance, sometimes comparable to supervised fine-tuning procedures, but at a fraction of the computational cost.

We present new algorithms for estimating and testing \emph{collision probability}, a fundamental measure of the spread of a discrete distribution that is widely used in many scientific fields. We describe an algorithm that satisfies $(\alpha, \beta)$-local differential privacy and estimates collision probability with error at most $\epsilon$ using $\tilde{O}\left(\frac{\log(1/\beta)}{\alpha^2 \epsilon^2}\right)$ samples for $\alpha \le 1$, which improves over previous work by a factor of $\frac{1}{\alpha^2}$. We also present a sequential testing algorithm for collision probability, which can distinguish between collision probability values that are separated by $\epsilon$ using $\tilde{O}(\frac{1}{\epsilon^2})$ samples, even when $\epsilon$ is unknown. Our algorithms have nearly the optimal sample complexity, and in experiments we show that they require significantly fewer samples than previous methods.

Constrained Markov decision processes (CMDPs) models are increasingly important in many applications with multiple objectives. When the model is unknown and must be learned online, it is desirable to ensure that the constraint is met, or at least the violation is bounded with time. In recent literature, progress has been made on this very challenging problem but with either unsatisfactory assumptions such as the knowledge of a safe policy, or have high cumulative regret. We propose the Safe-PSRL (posterior sampling-based RL) algorithm that does not need such assumptions and yet performs very well, both in terms of theoretical regret bounds as well as empirically. The algorithm efficiently trades-off exploration and exploitation using posterior sampling-based exploration, and yet provably suffers only bounded constraint violation using carefully-crafted pessimism. We establish a sub-linear $\tilde{O}(H^{2.5}\sqrt{|S|^2|A|K})$ upper bound on the Bayesian objective regret along with a bounded, i.e., $\tilde{O}(1)$ constraint-violation regret over $K$ episodes for an $|S|$-state, $|A|$-action, and $H$ horizon CMDP which improves over state-of-the-art algorithms for the same setting.

Group testing is the problem of identifying a small subset of defectives from a large set using as few binary tests as possible. In most current literature on group testing the binary test outcome is $1$ if the pool contains at least one defective, and $0$ otherwise. In this work we initiate the study of a generalized model of group testing that accommodates the physical effects of dilution of infected samples in large pools. In this model the binary test outcome is $1$ with probability $f(\rho)$, where $\rho$ is the density of the defectives in the test, and $f:[0,1]\rightarrow [0,1]$ is a given "test function" that models this dilution process. For a large class of test functions our results establish near-optimal sample complexity bounds, by providing information-theoretic lower bounds on the number of tests necessary to recover the set of defective items, and providing computationally efficient algorithms with sample complexities that match these lower bounds up to constant or logarithmic factors. Furthermore, using tools from real analysis, we extend our results to any "sufficiently well-behaved function" $f:[0,1]\rightarrow [0,1]$.

Local differential privacy (LDP) is increasingly employed in privacy-preserving machine learning to protect user data before sharing it with an untrusted aggregator. Most LDP methods assume that users possess only a single data record, which is a significant limitation since users often gather extensive datasets (e.g., images, text, time-series data) and frequently have access to public datasets. To address this limitation, we propose a locally private sampling framework that leverages both the private and public datasets of each user. Specifically, we assume each user has two distributions: $p$ and $q$ that represent their private and public datasets, respectively. The objective is to design a mechanism that generates a private sample approximating $p$ while simultaneously preserving $q$. We frame this objective as a minimax optimization problem using $f$-divergence as the utility measure. We fully characterize the minimax optimal mechanisms for general $f$-divergences provided that $p$ and $q$ are discrete distributions. Remarkably, we demonstrate that this optimal mechanism is universal across all $f$-divergences. Experiments validate the effectiveness of our minimax optimal mechanism compared to the state-of-the-art private sampler.

We study the canonical fair clustering problem where each cluster is constrained to have close to population-level representation of each group. Despite significant attention, the salient issue of having incomplete knowledge about the group membership of each point has been superficially addressed. In this paper, we consider a setting where the assigned group memberships are noisy. We introduce a simple noise model that requires a small number of parameters to be given by the decision maker. We then present an algorithm for fair clustering with provable \emph{robustness} guarantees. Our framework enables the decision maker to trade off between the robustness and the clustering quality. Unlike previous work, our algorithms are backed by worst-case theoretical guarantees. Finally, we empirically verify the performance of our algorithm on real world datasets and show its superior performance over existing baselines.

We study goal-conditioned Hierarchical Reinforcement Learning (HRL), where a high-level agent instructs sub-goals to a low-level agent. Under the assumption of a sparse reward function and known hierarchical decomposition, we propose a new algorithm to learn optimal hierarchical policies. Our algorithm takes a low-level policy as input and is flexible enough to work with a wide range of low-level policies. We show that when the algorithm that computes the low-level policy is optimistic and provably efficient, our HRL algorithm enjoys a regret bound which represents a significant improvement compared to previous results for HRL. Importantly, our regret upper bound highlights key characteristics of the hierarchical decomposition that guarantee that our hierarchical algorithm is more efficient than the best monolithic approach. We support our theoretical findings with experiments that underscore that our method consistently outperforms algorithms that ignore the hierarchical structure.

Retrosynthesis poses a key challenge in biopharmaceuticals, aiding chemists in finding appropriate reactant molecules for given product molecules. With reactants and products represented as 2D graphs, retrosynthesis constitutes a conditional graph-to-graph (G2G) generative task. Inspired by advancements in discrete diffusion models for graph generation, we aim to design a diffusion-based method to address this problem. However, integrating a diffusion-based G2G framework while retaining essential chemical reaction template information presents a notable challenge. Our key innovation involves a multi-stage diffusion process. We decompose the retrosynthesis procedure to first sample external groups from the dummy distribution given products, then generate external bonds to connect products and generated groups. Interestingly, this generation process mirrors the reverse of the widely adapted semi-template retrosynthesis workflow, i.e., from reaction center identification to synthon completion. Based on these designs, we introduce Retrosynthesis Diffusion (RetroDiff), a novel diffusion-based method for the retrosynthesis task. Experimental results demonstrate that RetroDiff surpasses all semi-template methods in accuracy, and outperforms template-based and template-free methods in large-scale scenarios and molecular validity, respectively.

This paper introduces posterior mean matching (PMM), a new method for generative modeling that is grounded in Bayesian inference. PMM uses conjugate pairs of distributions to model complex data of various modalities like images and text, offering a flexible alternative to existing methods like diffusion models. PMM models iteratively refine noisy approximations of the target distribution using updates from online Bayesian inference. PMM is flexible because its mechanics are based on general Bayesian models. We demonstrate this flexibility by developing specialized examples: a generative PMM model of real-valued data using the Normal-Normal model, a generative PMM model of count data using a Gamma-Poisson model, and a generative PMM model of discrete data using a Dirichlet-Categorical model. For the Normal-Normal PMM model, we establish a direct connection to diffusion models by showing that its continuous-time formulation converges to a stochastic differential equation (SDE). Additionally, for the Gamma-Poisson PMM, we derive a novel SDE driven by a Cox process, which is a significant departure from traditional Brownian motion-based generative models. PMMs achieve performance that is competitive with generative models for language modeling and image generation.

Several generative models with elaborate training and sampling procedures have been proposed to accelerate structure-based drug design (SBDD); however, their empirical performance turns out to be suboptimal. We seek to better understand this phenomenon from both theoretical and empirical perspectives. Since most of these models apply graph neural networks (GNNs), one may suspect that they inherit the representational limitations of GNNs. We analyze this aspect, establishing the first such results for protein-ligand complexes. A plausible counterview may attribute the underperformance of these models to their excessive parameterizations, inducing expressivity at the expense of generalization. We investigate this possibility with a simple metric-aware approach that learns an economical surrogate for affinity to infer an unlabelled molecular graph and optimizes for labels conditioned on this graph and molecular properties. The resulting model achieves state-of-the-art results using 100x fewer trainable parameters and affords up to 1000x speedup. Collectively, our findings underscore the need to reassess and redirect the existing paradigm and efforts for SBDD. Code is available at https://github.com/rafalkarczewski/SimpleSBDD.

This work studies the generalization error of gradient methods. More specifically, we focus on how training steps $T$ and step-size $\eta$ might affect generalization in smooth stochastic convex optimization (SCO) problems. Recent works show that in some cases longer training can hurt generalization. Our work reexamines this for smooth SCO and find that the conclusion can be case-dependent. In particular, we first study SCO problems when the loss is \emph{realizable}, i.e. an optimal solution minimizes all the data points. Our work provides excess risk lower bounds for Gradient Descent (GD) and Stochastic Gradient Descent (SGD) and finds that longer training may not hurt generalization. In the short training scenario $\eta T = O(n)$ ($n$ is sample size), our lower bounds tightly match and certify the respective upper bounds. However, for the long training scenario where $\eta T =O(n)$, our analysis reveals a gap between the lower and upper bounds, indicating that longer training does hurt generalization for realizable objectives. A conjecture is proposed that the gap can be closed by improving upper bounds, supported by analyses in two special instances. Moreover, besides the realizable setup, we also provide first tight excess risk lower bounds for GD and SGD under the general non-realizable smooth SCO setting, suggesting that existing stability analyses are tight in step-size and iteration dependence, and that overfitting provably happens when there is no interpolating minimum.

We investigate how to select the number of communities for weighted networks without a full likelihood modeling. First, we propose a novel weighted degree-corrected stochastic block model (DCSBM), where the mean adjacency matrix is modeled in the same way as in the standard DCSBM, while the variance profile matrix is assumed to be related to the mean adjacency matrix through a given variance function. Our method of selecting the number of communities is based on a sequential testing framework. In each step, the weighted DCSBM is fitted via some spectral clustering method. A key component of our method is matrix scaling on the estimated variance profile matrix. The resulting scaling factors can be used to normalize the adjacency matrix, from which the test statistic is then obtained. Under mild conditions on the weighted DCSBM, our proposed procedure is shown to be consistent in estimating the true number of communities. Numerical experiments on both simulated and real-world network data demonstrate the desirable empirical properties of our method.

In interactive systems, actions are often correlated, presenting an opportunity for more sample-efficient off-policy evaluation (OPE) and learning (OPL) in large action spaces. We introduce a unified Bayesian framework to capture these correlations through structured and informative priors. In this framework, we propose sDM, a generic Bayesian approach for OPE and OPL, grounded in both algorithmic and theoretical foundations. Notably, sDM leverages action correlations without compromising computational efficiency. Moreover, inspired by online Bayesian bandits, we introduce Bayesian metrics that assess the average performance of algorithms across multiple problem instances, deviating from the conventional worst-case assessments. We analyze sDM in OPE and OPL, highlighting the benefits of leveraging action correlations. Empirical evidence showcases the strong performance of sDM.