🏢 New York University

Unlocking tight generalization bounds for massive LLMs using a novel token-level approach.

TuneTables optimizes PFNs for scalability via context optimization, achieving state-of-the-art performance on large tabular datasets while using fewer parameters and reducing inference time.

Optimization’s implicit bias in robust machine learning hurts generalization; this work reveals how algorithm/architecture choices impact robustness, suggesting better optimization strategies are need…

Reinforcement learning achieves unprecedented fast convergence rates in continuous state-action spaces by leveraging novel stability properties of Markov Decision Processes.

Contextual bandits with graph feedback achieve near-optimal regret by leveraging a novel graph-theoretic quantity that interpolates between independence and maximum acyclic subgraph numbers, depending…

Revolutionizing large neural networks, this paper introduces a continuous parameterization of structured matrices, discovering that full-rank structures without parameter sharing achieve optimal scali…

Overparameterized neural networks surprisingly recover from catastrophic interference when trained cyclically on repeated data sequences, exhibiting anticipatory knowledge reactivation.

Boosting posterior sampling in challenging high-dimensional inverse problems, this paper introduces ’tilted transport’, a novel technique leveraging denoising oracles for provably easier sampling.

Despite common belief, state-of-the-art preference learning algorithms for LLMs achieve surprisingly low ranking accuracy, highlighting significant flaws in current alignment techniques.

HOAM learns reduced models of population dynamics for complex systems, enabling fast predictions across various physics parameters, outperforming state-of-the-art techniques.

RUM neural network, a novel non-convolutional GNN, overcomes limitations of conventional convolution-based models by using RNNs to merge topological and semantic features along random walks, achieving…

Sparse navigable graphs enable efficient nearest neighbor search, but their construction and limits in high dimensions remain unclear. This paper presents an efficient method to construct navigable gr…

AI models struggle to understand 3D space like humans do. This paper introduces Multiview Scene Graphs (MSGs) – a new topological scene representation using interconnected place and object nodes buil…

This paper introduces robust Dikin walks for log-concave sampling, achieving faster mixing times and lower iteration costs than existing methods, particularly for high-dimensional settings.

Teach LLMs uncertainty for reliable high-stakes predictions: Fine-tuning with graded examples significantly improves LLM’s uncertainty calibration and generalizes well.

New method, STRIPE-X, powerfully detects ’encoding’ in AI explanations—a sneaky phenomenon where explanations predict outcomes better than their constituent parts alone would suggest.

Faster, more efficient deep learning for diffusion MRI deconvolution is achieved using spatio-hemispherical networks, improving fiber tractography.

Prompt Gradient Alignment (PGA) enhances unsupervised domain adaptation by aligning per-objective gradients in a multi-objective optimization framework, achieving state-of-the-art results.

DynaMo: a novel self-supervised method significantly boosts visuo-motor control by learning in-domain dynamics from limited expert demonstrations, improving policy performance across various environme…

Symile: A simple model-agnostic approach for learning representations from unlimited modalities, outperforming pairwise CLIP by capturing higher-order information.