All papers

The paper introduces RefineMix to train discrete diffusion models under data scarcity using out-of-distribution data at chosen diffusion timesteps.

The paper presents SplitJEPA, learning invariant and variant latent world state representations directly without using reconstruction objectives.

This paper derives sharp sample complexity thresholds for Gaussian regression over bounded-degree polynomial subspaces on the Boolean cube.

This paper establishes tighter regret bounds for adversarial Markov decision processes by incorporating regularized Q-functions into policy optimization.

This paper uses bilevel optimization and kernel autoencoders to learn finite-dimensional Koopman operator embeddings from nonlinear dynamical data.

This paper presents a transformer architecture that predicts conditional marginal distributions for missing data in a single forward pass.

This paper introduces a regional weather forecasting framework that aligns pretrained global model guidance with high-resolution regional representations.

This paper provides statistical sample complexity guarantees for transformer-based masked diffusion models learning distributions with latent tree structures.

This paper investigates the relationship between circuit depth, training dataset size, and generalization bounds in adaptively grown quantum classifiers.

This paper trains a generalist humanoid policy by predicting body and hand motion latents mapped from human demonstration data.

This paper estimates density ratios by parameterizing distribution shift as a displacement field using the Stein operator.

This paper redesigns 4-bit AdamW optimizer-state quantization by performing stochastic rounding directly within preconditioner space.

This paper analyzes data sampling distributions across parallel simulations to improve exploration efficiency in reinforcement learning for robot control.

This paper proposes a sample-adaptive zeroth-order optimization method that estimates sample importance online to accelerate large language model fine-tuning.

This paper designs and simulates an all-transistor physical circuit using sub-threshold MOSFETs to run transformer attention and in-context learning.

This paper proposes a reward aggregation method to stabilize and scale reinforcement learning fine-tuning for diffusion models.

This paper introduces a benchmark and a diffusion aggregation algorithm to evaluate and balance accuracy and realism in event sequence forecasting.

This paper presents a method for upcycling pretrained language models into sparse architectures using token-indexed embeddings in residual pathways.

This paper applies a convolutional Gaussian neural process and genetic optimization to select optimal aerosol monitoring sensor locations across the UK.

This paper introduces a one-step implicit sampler using short-time MCMC dynamics to enable efficient amortized inference for Bayesian posteriors.

This paper uses a ConvNeXt architecture to detect low-voltage EEG intervals as biomarkers for neonatal hypoxic-ischaemic encephalopathy.

This paper optimizes the Maximum Common Edge Subgraph distance calculation to evaluate structure-disjoint data splits on large molecular libraries.

Combines structural graph backbone construction with reinforcement learning completion to design network topologies that maximize algebraic connectivity.

Introduces primal-dual flow matching to enforce sample-wise constraints in generative modeling using binary membership feedback.

Proposes Product Monte Carlo to compute expectations under products of densities using draws from individual densities via regression.

Analyzes the optimal balance between spatial sensor window size and temporal history when forecasting partially observed physical fields.

Develops a category-theoretic framework to analyze and compare parameter-efficient fine-tuning methods as smooth maps in weight space.

Establishes minimax estimation rates and examines plug-in estimators for powered sliced Wasserstein functionals under moment bounds.

Interprets Transformer networks as learned approximations to loopy belief propagation and expectation-maximization in dynamic Markov random fields.

Reports that initialization-time gradient masks underperform random masks on catastrophic forgetting when confined to the same LoRA parameter pool.

Introduces a dual-curriculum conditional diffusion framework for multivariate time-series anomaly detection across varying masking and noise levels.

Proposes an institutional provenance-grouped evaluation protocol to eliminate data leakage when benchmarking models for perovskite photovoltaics.

Applies multivariate Hawkes processes to sequence data of soccer passes and shots to distribute attacking credit across multiple players.

Presents a coordinate-free automatic differentiation formulation that expresses geometric differential operations within standard machine learning primitives.

Develops polyconvex neural network surrogate models using learned paired directional features for anisotropic hyperelastic material homogenization.

Investigates quantum-native byte-level language models, analyzing memory limits under learned dissipation and testing them on quantum hardware.

Analyzes the expressive power and size-uniform approximation limits of message-passing neural networks using continuous random node features.

Presents deployment experiences and design strategies for ByteExpress, an RDMA communication layer for distributed training and inference workloads.

Uses ensembles of conditional variational autoencoders with wavelet decoders to synthesize acceleration time series from shock response spectra.

Develops a tree-based exploration and hindsight credit assignment method to improve multi-turn reinforcement learning for language model agents.

Introduces a scalable training method for determinantal point processes using nonlinear kernels, latent features, and stochastic optimization.

Proposes a cross-behavior policy optimization framework unifying offline demonstrations and online rollouts via a shared regression objective.

Presents fixed-point flow matching, a data-free procedure that trains flow policies in reinforcement learning without requiring demonstration data.

Introduces a two-way guidance framework that jointly learns sufficient summary statistics and posteriors for scalable simulation-based inference.

Establishes finite-sample bounds for self-normalized Kullback-Leibler divergence estimators and applies them to test-time alignment sampling techniques.

Analyzes Langevin sampling algorithms under state-dependent heavy-tailed gradient noise with applications to Bayesian fine-tuning of language models.

Proposes a weighted metric to quantify and predict performance variability across mixed-integer programming solver runs.

Introduces a hybrid attention-RWKV architecture to mitigate horizon degradation in non-autoregressive time series forecasting.

Presents a cooperative game theoretic semi-supervised regression framework for predicting reservoir properties from seismic data.

Introduces determinant-ratio point processes to jointly model attraction and repulsion with distinct symmetric positive-definite kernels.

Proposes averaging query losses over generator states to improve pretraining sample efficiency in tabular foundation models.

Applies continuous-time control theory to stabilize adversarial minimax optimization against rotational divergence.

Analyzes online binary logistic regression where features are observed noisily subject to variable acquisition costs.

Studies dictionary learning on probability measures via linearized optimal transport embeddings and sparse autoencoders.

Proposes an imbalance-aware deep nonparametric two-sample test that leverages majority observations through repeated testing evidence.

Develops a matching repair framework with theoretical performance guarantees for training data acquisition under delayed feedback.

Presents a distributed system architecture aggregating idle Apple Silicon devices into a virtual GPU cloud for inference.

Employs a conditional diffusion model to generate observation-aligned precipitation forecasts from frozen weather foundation models.

Introduces a nuclear-aware determinant ansatz for multi-geometry neural-network variational Monte Carlo calculations.

Proposes scan-parallel variational inference for continuous-time latent jump models with state-dependent event intensities.