All papers

The authors present an autotuning framework for WebGPU kernels to improve cross-platform inference speeds for language models on edge devices.

The authors present an ensemble variational data assimilation algorithm that incorporates learned generative forecasting models to account for complex model errors.

The authors introduce a parameter-free conformal martingale method modeled as a universal portfolio to quickly detect distribution shifts at unknown times.

The authors propose a test-time learning method that trains a proposer to generate prompts, tools, and controls guiding scientific discovery agents.

The authors analyze robustly reliable learning for regression under data corruptions and propose an algorithm to compute optimal prediction intervals for Lipschitz predictors.

The authors introduce a masked autoencoder that pretrains wearable sensor representations across varied recording durations and sampling rates using cross-resolution views.

The authors define measure neural mappings and establish quantitative error bounds for self-attention transformers acting on probability measures in unbounded spaces.

The authors present a mathematical method to enforce Dirichlet, Neumann, and Robin boundary conditions exactly on curved triangular domains for scientific machine learning.

The authors introduce an on-policy distillation method to transfer knowledge from autoregressive language models into diffusion language models using autoregressive rollouts.

The authors argue that cooperative multi-agent reinforcement learning systems require test-time coordination mechanisms rather than relying solely on training-time centralized information.

The authors develop a reachability-guided dynamic sparse training method that preserves functional pathways during extreme network pruning and regrowth.

The authors propose an optimizer for Monarch-factored weights that replaces dense matrix orthogonalization with smaller block updates to reduce memory and pretraining time.

The authors present a verification workflow combining reachability analysis with Bayesian optimization to close residual test bins in chip design verification.

The authors compare self-supervised and sleep-stage-supervised representations for identifying sleep-wake disorders across thousands of polysomnography recordings.

The authors analyze continual learning on a spherical parameter space to determine how many tasks a model can learn without catastrophic forgetting.

The authors introduce a selective hypothesis test for comparing cluster means under an unknown covariance matrix without inflating false-positive rates.

The paper proposes a hierarchical reinforcement learning framework to automate multi-phase spacecraft rendezvous and docking trajectories.

The paper introduces diagnostic scores to test whether neural network surrogates respect underlying physical conservation laws rather than numerical errors.

The paper evaluates online machine learning models within a time-synchronous framework to predict weekly student disengagement from activity logs.

The paper evaluates clustering methods on learning management system logs to identify learner participation profiles across unequal observation windows.

The paper introduces a multimodal auditory attention decoding framework designed to maintain reliance on neural signals alongside behavioural tracking.

The paper presents an automated coding agent that tunes large model training configurations for throughput under user-defined compute budgets.

The paper proposes an inference-time alignment method that optimizes expected costs while controlling tail risk in diffusion models.

The paper introduces a self-distillation framework where a vision-language model teaches itself language-conditioned robotic navigation policies without human annotations.

The paper introduces a personalized federated learning method that regularizes local model updates using low-rank difference structures.

The paper proposes mixed precision scaling laws to model the trade-offs of low-precision quantization during model pretraining.

The paper analyzes the theoretical detection limits and sum-test efficiency of textGrain watermarking under human text edits.

The paper introduces a global-to-local dynamic attention model to filter visual noise and improve multimodal sarcasm detection.

The paper proposes a concept-guided regularizer using sparse autoencoders to reduce catastrophic forgetting during parameter-efficient fine-tuning.

The paper develops a flow matching framework with structured priors to model partially observed time series without wasting capacity on observed entries.

The paper investigates the representation geometry and inductive biases of quantum classifiers when applied to classical visual benchmark datasets.

The paper proposes a Kolmogorov-Arnold Network architecture using parallel Bessel branches and quadratic interactions to reduce parameters and conditioning problems.

The paper provides an adversarial linear bandit algorithm achieving sublinear regret over general convex action sets using follow-the-perturbed-leader.

The paper presents a geometry-aware pretraining framework that aligns heterogeneous probe layouts for functional near-infrared spectroscopy representation learning.

The paper proposes a minimal positional encoding method using random Fourier features to feed persistence diagrams directly into standard Transformers.

The paper develops a neural network approach to learn effective mesoscopic field equations directly from microscopic trajectory data.

The paper presents an expressive and distributed attention kernel that compresses structured attention masks to enable efficient context parallelism on Hopper GPUs.

The paper presents a subspace construction method for continual low-rank adaptation to mitigate forgetting under differential privacy constraints.

The paper introduces a unified local training framework using norm-ball clipping to improve communication efficiency in distributed and federated learning.

The paper introduces a score-guided particle framework using Stein variational gradient descent to model multivariate return distributions in reinforcement learning.

This paper establishes accelerated parameter estimation rates when continuous data samples are accompanied by binary quality labels or pairwise preferences.

This paper proposes a conformal risk control layer that certifies miss rates and reduces alarm duty cycles in predictive cellular network operations.

This paper introduces a dimensional analysis framework and diagnostic metrics to maintain scale consistency across Transformer components.

This paper presents an error-controlled compression codec for key-value caches using coordinate alignment and predictive coding.

This paper presents an equivariant neural network functional to perform orbital-free density functional theory for periodic CHNO systems.

This paper introduces an exact method to compute Banzhaf and Shapley power indices for base classifiers in plurality-voting ensembles.

This paper proposes a counterfactual certification framework to bound performance loss during sequential machine unlearning requests.

This paper improves neural symbolic regression by using a tabular prior and reinforcement learning alignment to fit noisy data.

This paper theoretically and empirically compares weighting-based rebalancing strategies to unweighted methods in imbalanced tabular classification tasks.

This paper introduces a robot manipulation benchmark to isolate failures across scaling subgoal horizons and orchestration methods.

This paper develops an empirical-likelihood bandit policy that provides asymptotic regret guarantees without knowledge of reward scales.

This paper proposes a score-matching approach for causal discovery in additive noise models under hard selection bias and truncation.

This paper introduces a framework combining persistent belief states with implicit world modeling to help long-horizon language agents recover from failures.

This paper demonstrates that standard optimizers are sensitive to parameter scaling in Transformers and proposes a symmetry-aware gradient adjustment.

This paper analyzes prediction errors in spectral saliency when deleting singular components to selectively edit language model capabilities.

This paper analyzes how layer ordering and selection affect performance across tasks in hybrid attention model architectures.

This paper introduces an attention-based neural network to predict code distances for bivariate bicycle quantum low-density parity-check codes.

This paper formulates a trace-minimization framework to design causal error-feedback coefficients for neural network weight quantization.

This paper analyzes the sample and mistake complexity of learning group elements from actions using Galois correspondences and multiclass dimensions.

This paper proposes a message-passing layer using randomized probing of operator diagonals to capture node self-interactions in graph neural networks.