AI Theory

Cost-aware Bayesian optimization gets a boost with the Pandora’s Box Gittins Index, a novel acquisition function that efficiently balances exploration and exploitation while considering evaluation cos…

This paper presents novel algorithms for linear bandits that are robust to corrupted rewards, achieving minimax optimality and optimal scaling for gap-dependent misspecification, extending to reinforc…

In symmetric zero-sum games, no-swap-regret dynamics guarantee strong convergence to Nash Equilibrium under symmetric initial conditions, but this advantage disappears when constraints are relaxed.

Gradient-based methods for solving large zero-sum games achieve polynomial smoothed complexity, demonstrating efficiency even in high-precision scenarios without condition number dependence.

New PAC-Bayes bound controls multiple error types simultaneously, providing richer generalization guarantees.

AI decision support systems can unintentionally harm users; this paper introduces a novel framework to design systems that minimize this counterfactual harm, balancing accuracy and user well-being.

Contrastive losses unlock efficient fitness function modeling by leveraging the ranking information inherent in global epistasis, significantly improving accuracy and data efficiency in protein engine…

Repeated contracts with learning agents are optimized by a simple dynamic contract: initially linear, then switching to zero-cost, causing the agent’s actions to ‘free-fall’ and yield non-zero rewards…

Neural networks, viewed as Bayesian ensembles of fixed classifiers, enable continual learning without forgetting; posterior updates mirror stochastic gradient descent, offering insights into optimizat…

Continual counting with gradual privacy expiration: A new algorithm achieves optimal accuracy with exponentially decaying privacy!

This paper introduces induced empirical risk minimization for contextual linear optimization with bandit feedback, providing theoretical guarantees and computationally tractable solutions for improved…

This work unveils a novel algorithm for contextual decision-making with knapsacks, achieving significantly improved regret bounds beyond worst-case scenarios, thereby offering a more practical and eff…

Constrained sampling made easy! Primal-Dual Langevin Monte Carlo efficiently samples from complex probability distributions while satisfying statistical constraints.

This paper presents a unified framework for solving binary statistical decision-making problems, enabling efficient derivation of optimal strategies for diverse applications like OOD detection and sel…

Constrained Adaptive Attack (CAA) significantly improves adversarial attacks on deep learning models for tabular data by combining gradient and search-based methods, achieving up to 96.1% accuracy dro…

Neural causal models consistently estimate partial causal effects, even with continuous/categorical variables, thanks to Lipschitz regularization.

Conformal inverse optimization learns uncertainty sets for parameters in optimization models, then solves a robust optimization model for high-quality, human-aligned decisions.

This paper introduces AFCP, a novel conformal inference method that generates prediction sets with valid coverage conditional on adaptively selected features, achieving a practical balance between eff…

Researchers introduce the Conditional Quantile Comparator (CQC) for analyzing heterogeneous treatment effects, offering an improved approach by combining the strengths of CATE and CQTE while overcomin…

ID-GEN: Sample high-dimensional interventional distributions using any conditional generative model!