AI Theory

MILP-StuDio generates high-quality mixed-integer linear programming instances by preserving crucial block structures, significantly improving learning-based solver performance.

MG-Net dynamically designs optimal mixer Hamiltonians for QAOA, overcoming the limitation of fixed-depth quantum circuits and significantly improving approximation ratios.

Researchers unveil novel linear-algebraic tools revealing the limits of fast attention, classifying positive definite kernels for Manhattan distance, and fully characterizing metric transforms for Man…

Novel metric space magnitude measures rigorously quantify the diversity of latent representations across multiple scales, showing superior performance in detecting mode collapse and characterizing emb…

Mechanism design enhanced with output advice improves approximation guarantees by using imperfect predictions of the output, not agent types, offering robust, practical solutions.

New metric, Maximum Entropy Goal-Directedness (MEG), quantifies AI goal-directedness, crucial for assessing AI safety and agency.

This paper presents a novel bilevel approach to extend mean-field Langevin dynamics to solve convex optimization problems over signed measures, achieving stronger guarantees and faster convergence rat…

Researchers establish basis-free conditions for SGD learnability in two-layer neural networks learning subspace-sparse polynomials with Gaussian input, offering insights into training dynamics.

Optimizing against learning agents: New algorithms and computational limits revealed!

MatrixNet learns efficient group representations for improved deep learning on symmetry groups, achieving higher sample efficiency and generalization than existing methods.

Optimal matrix denoising with doubly heteroscedastic noise achieved!

Sign Stochastic Gradient Descent (SGD) achieves optimal sample complexity for solving k-sparse parity problems, matching Statistical Query lower bounds.

Masked hard-attention transformers, with strict masking, precisely capture star-free languages, matching the expressive power of linear temporal logic.

This study marries causal representation learning with dynamical systems to enable parameter identification in real-world scientific data, unlocking downstream causal analysis for various applications…

Researchers developed a new, differentiable score function for learning causal relationships from data that reliably recovers the simplest causal model, even with complex data.

Frugal Flows: Generate realistic causal benchmarks with exact marginal causal effects, enabling robust causal method validation.

MambaLRP enhances explainability of Mamba sequence models by ensuring faithful relevance propagation, achieving state-of-the-art explanation performance, and uncovering model biases.

MALT: a novel adversarial attack, is 5x faster than AutoAttack, achieving higher success rates on CIFAR-100 and ImageNet by exploiting mesoscopic almost linearity in neural networks.

Improved facility location mechanisms are designed using ‘Mostly Approximately Correct’ predictions, exceeding prior bounds despite large prediction errors.

Low-degree polynomials fail to efficiently infer roots in broadcasting tree problems below the Kesten-Stigum bound.