All papers

Introduces continuous-time inter-variable attention inside differential equation solvers for long-term multivariate time-series forecasting.

Models reward overoptimization in reinforcement learning with human feedback as economic rent-seeking driven by proxy reward signals.

Benchmarks the inference energy efficiency of diffusion language models compared to standard autoregressive architectures.

Applies large language models to filter anomalies in industrial sensor training data using unstructured operational logs.

Analyzes the optimization landscape of overparameterized weighted maximum mean discrepancy to provide finite-particle guarantees.

Mitigates catastrophic forgetting in continual learning through a hybrid approach combining parameter isolation and local regularization.

Proposes a one-step generative approach to model occupancy for simplified policy extraction in reinforcement learning.

Improves industrial sound anomaly detection by adapting unsupervised models with few fault samples to reduce false alarms.

Predicts high-level electronic band structures of materials using orbital information combined with low-cost baseline calculations.

Establishes theoretical convergence bounds for Gaussian process Thompson sampling under noiseless conditions in global optimization.

Treats symbolic regression as a continuous optimization problem using neural units that morph dynamically into mathematical formulas.

Allocates computation unevenly across passes in looped mixture-of-experts models to improve efficiency without losing performance.

The paper develops a minimax-optimal estimator for the pseudo-spectral gap of continuous-state Markov chains from a single observed trajectory.

The paper demonstrates that hyperbolic contrastive reinforcement learning achieves strong navigation performance with much shallower networks than Euclidean models.

The paper introduces a time-dependent control-cost weighting method to fine-tune diffusion and flow-matching models with reward guidance.

The paper proposes a staged physics-informed neural network to estimate electrical parameters of a direct-current buck power converter.

The paper develops twin support vector regression methods using granular ball computing to improve robustness against noise and outliers.

The paper uses random matrix theory to analyze the trade-off between noise reduction and bias in decentralized federated learning.

The paper evaluates autoregressive foundation models trained on simulated market order flows to analyze whether they learn underlying Bayesian dynamics.

The paper introduces an exhaustive dataset of univariate expression skeletons generated by a formal grammar to improve symbolic regression.

The paper presents an interactive, browser-based tool for visualizing and learning the mechanics of Kolmogorov-Arnold Networks.

The paper introduces a Bregman proximal framework for neural operators to interpret activation functions and improve the training of deeper models.

The paper measures the energy consumption, latency, and thermal behavior of running vision-language models on edge hardware accelerators.

The paper proposes an amortized flow-matching method to generate diverse batches for parallel black-box optimization over sequential rounds.

The paper presents a mixed integer linear programming formulation to train sparse, interpretable Boolean rule networks with reduced decision variables.

The paper derives dimension-optimal dynamic regret bounds for online learning with exp-concave losses under time-varying comparator paths.

The paper improves behavior cloning for autonomous driving using bird's-eye view inputs and action-density weighting for rare events like traffic lights.

The paper proposes a heterogeneous graph neural network that combines circuit topology, physical layout, and routing resources to predict chip routability.

The paper introduces diffusion models driven by Poisson counting processes to solve discrete Poisson inverse problems via posterior sampling.

The paper uses differential topology to establish theoretical lower bounds on the minimal width of universal approximator neural networks.

The paper presents a replay-free dynamic mixture-of-experts method to mitigate recency bias in class-incremental graph learning without needing task identifiers.

The paper uses adversarial reinforcement learning with physically constrained perturbations to improve the robustness of autonomous planetary landing controllers.

The paper introduces a rectified flow matching framework to separate overlapping radio-frequency signals in a single channel in one step.

The paper presents a multi-agent reinforcement learning framework for designing RNA sequences by co-optimizing both secondary topology and tertiary geometry.

The paper introduces a lattice-based theoretical framework to reconcile differing hypotheses about how internal representations converge across neural networks.

The paper investigates how conditional flow models learn and potentially hallucinate non-additive interactions in scientific prediction tasks.

This paper proposes a closed-form covariance-aligned rotation method to improve post-training low-bit quantization in large language models.

This paper explores extreme super-resolution of turbulent flow fields by training a deep neural network on a single flow realization.

This paper presents an entropy-regularized memory-guided policy optimization method for online test-time adaptation in robot locomotion.

This paper evaluates classical and deep learning direction-of-arrival estimators across accuracy and latency metrics for real-time integrated sensing and communication.

This paper introduces a two-stage priming framework combining reconstruction pretraining with distributional post-training for one-step generative modeling.

This paper proves that training shallow transformers on unbounded inputs causes attention weight learning to lag behind output layer learning.

This paper presents a task-conditioned imputation framework and evaluates model robustness under train-test missingness mechanism mismatches.

This paper proposes a discrete spatiotemporal tokenization framework for asynchronous event detection and boundary estimation in continuous EEG recordings.

This paper proposes a contrast-calibrated bit allocation method for mixed-precision quantization of mixture-of-experts language models.

This paper shows that providing learnable fractional derivative features improves theoretical bounds and empirical accuracy in neural operator learning.

This paper introduces a flow-matching plug-in module to recover high-frequency physical structures in pretrained neural partial differential equation solvers.

This paper provides a fast hypothesis selection algorithm using weighted medians that improves runtime dependence on failure probability.

This paper presents a mathematical review analyzing when human corrections translate into reusable learning versus interactive refinements in image segmentation.

This paper develops a neural operator framework to upsample and personalize continuous head-related transfer function fields from sparse spatial measurements.

This paper introduces a reinforcement learning approach that trains language models to execute safe fallbacks instead of unreliable guesses.

This paper provides counterexamples to an established memorization lower bound and derives exact replacements using full-sample posterior probabilities.

This paper presents a self-supervised reinforcement learning method that uses hyperspherical representations and spherical interpolation to infer planning waypoints.

This paper proves a computational separation between information-theoretic and polynomial-time regret when tracking small expert subsets in online convex optimization.

This paper introduces a continual learning replay method to incrementally evaluate and adapt molecular communication models across changing experimental settings.

This paper analyzes non-monotone accuracy dynamics during early neural network training when models are initialized with vanishingly small weights.

This paper evaluates whether a pretrained tabular foundation model can perform power grid steady-state analysis without task-specific retraining.

This paper proposes a decoupled marginal sharpening method to improve inference-time reasoning in language models without weight updates or reward models.

This paper provides matching upper and lower bounds for learning near-optimal policies from a single continuing trajectory in average-reward reinforcement learning.

This paper introduces a graph-enhanced framework to discover disease trajectories and forecast temporal clinical endpoints from electronic health records.