All papers

The paper studies query complexity bounds and limits of momentum-based acceleration in Gaussian variational inference for strongly log-concave distributions.

The paper provides theoretical last-iterate convergence guarantees for clipped stochastic gradient methods under heavy-tailed noise.

The paper introduces an operator-learning framework using linear transfer operators to infer localized finite-amplitude source events in nonlinear dynamical systems.

The paper provides theoretical bounds on the approximation error caused by score mismatch during zero-shot conditional sampling in discrete diffusion models.

The paper introduces a two-stage adaptive feature construction method to improve randomized neural networks for solving partial differential equations.

The paper proposes a heavy-tailed score-matching framework for limit order book trend prediction without needing reverse diffusion sampling at inference.

The paper proposes Noise Policy Optimization to align few-step generative models with black-box rewards by steering their initial input noise space.

The paper presents an attention-based neural network and logistic regression ensemble to predict live birth outcomes before fresh embryo transfer.

The paper analyzes theoretical properties of the multiplicative half-Cauchy process and provides a calibration rule linking prior beliefs to its scale parameter.

The paper proposes a rank-local data-loading policy that rebalances variable-cost training batches without cross-node coordination to speed up distributed training.

The paper evaluates two machine learning procedures to diagnose faulty sensors and correct fuel and emission estimates in heavy-duty diesel engines.

The paper introduces a training-free weight-editing method to reduce large language model compliance with harmful instructions such as lying and sexism.

The paper investigates data poisoning attacks targeting pre-training and retrieval corpora of protein language models and generative design systems.

The paper provides a high-dimensional theoretical analysis of data pruning strategies when training models on the hardest examples.

The authors introduce a graph neural ordinary differential equation framework to forecast multiscale spatiotemporal epidemic spreading dynamics.

The paper introduces an update-matched evaluation reference to balance unlearning effectiveness with downstream learning ability in language models.

The authors propose a theoretical and empirical framework to route token-level teacher selection during on-policy student distillation.

The paper presents a moment-matching framework for weakly supervised learning when labels are grouped into bags non-randomly.

The authors establish regret lower bounds for online quantum state learning under logarithmic loss with applications to tomography.

The study investigates whether removing small subsets of attention heads can improve language model performance on specific tasks.

The paper analyzes personalized projected ascent algorithms for federated minimax problems with client-specific maximization variables.

The authors use random matrix theory to establish theoretical limits on synthetic data-free preconditioning in differentially private learning.

The paper develops dimension-free PAC-Bayesian generalization bounds for deep neural networks using Dirichlet processes over data distributions.

The authors introduce online algorithms to generate quantile forecasts that remain accurate across multiple proper loss functions simultaneously.

The paper models demand management using probabilistic service options and analyzes provider incentives under transparency constraints.

The paper introduces a regularization technique that aligns internal representations across rollout origins to improve neural differential equation emulation.

The authors show how prequential regret from minimum description length bounds neural network generalization error without held-out data.

The authors implement an on-device machine learning diagnostic engine on a microcontroller for optical wireless communication links.

The paper proposes a graph neural network framework to exploit multiple spatiotemporal aggregation hierarchies for time-series forecasting.

This study audits how historical information and calendar controls affect trajectory forecasting within a university simulator.

The paper presents a predictive verifier that decides how many actions from a vision-language-action chunk to execute safely.

The authors propose an algorithm that assigns multi-reward reinforcement learning credit specifically to relevant response tokens.

The paper introduces a dual-encoder transformer that jointly transcribes piano performances to scores and tracks metrical beat frames.

This paper analyzes Kolmogorov-Arnold networks to show that favorable dimension scaling depends on target smoothness rather than the architecture alone.

The authors introduce a nonlinear phase kernel to expand the effective retrieval capacity of modern Hopfield networks.

The paper analyzes performance bottlenecks when using remote key-value caches during context-parallel prefill in large language model serving.

The paper analyzes the mathematical approximation capabilities of neural networks when layer normalization or RMSNorm is placed directly adjacent to activations.

The paper develops a primal-dual reinforcement learning algorithm providing finite-time performance and constraint-satisfaction bounds under partial observability.

The paper presents a differentiable masking technique to learn sparse control phases for preparing bosonic qudit quantum states.

The paper applies quantum natural language processing circuits to content-based TV programme recommendations, comparing efficiency against classical models.

The paper derives mathematical bounds to evaluate the calibration of query-based probabilistic models using logical consistency without ground-truth labels.

The paper establishes regret bounds and an algorithm for transfer learning in multi-armed bandits when historical data comes from unknown adaptive policies.

The paper applies game theory to analyze market incentives for AI safety investments across upstream foundation model providers and downstream firms.

The paper uses local temporal contrastive learning on human motion data to preserve multi-scale latent geometric trajectories in representations.

The paper presents a replay-driven decoupled reinforcement learning framework to improve system efficiency during language model post-training.

The paper proposes conditional mean flows to enable few-step posterior sampling and tractable density evaluation for simulation-based inference.

The paper analyzes the accuracy and efficiency trade-offs of derivative-free weight perturbation compared with standard backpropagation on MNIST.

The paper presents a lightweight classical latent variable method for generating synthetic mixed tabular data without complex deep architectures.

The paper proposes an interpretable, momentum-aware adaptive testing system that updates learner proficiency ratings dynamically during timed assessments.

The paper details a quantization-aware training workflow that trains neural networks to match the exact hardware constraints of edge accelerators.

The paper introduces a stochastic approximation expectation maximization algorithm for robust covariance estimation with coordinate-wise contaminated and missing data.

The paper presents a sparse federated Echo State Network using smooth L1 regularization for efficient edge time-series classification.

The paper analyzes direct pairwise area-under-curve training to explain why mini-batch stochastic gradient descent can fail on severe ranking errors.

The paper introduces a multi-horizon reinforcement learning framework that dynamically mixes multiple discount factors using policy confidence signals.

The paper investigates optimization dynamics in variational quantum circuits, showing single-qubit re-uploading circuits encounter phase-locking during first-order training.

The paper provides analytical formulas to characterize and explain the gradient of learning-rate transfer curves across different model scales.

The paper derives a closed-form geometric law describing how depth and width affect representation compression in deep linear networks.

The paper proposes an ONNX compiler method to map quantized non-monotonic and fine-granular activations into multi-threshold operators for hardware accelerators.

The paper frames client local training in federated learning as an optimal stopping problem to balance local updates against drift.

The paper distils a power-grid foundation model into a compact student network for real-time edge inference on power systems.