All papers

This paper introduces a benchmark and dataset that pairs numerical time-series data with real-world event context for forecasting.

This paper introduces a pre-trained vision transformer that combines experimental electron-density maps with atomic coordinates to model protein-ligand interactions.

This paper introduces a variational autoencoder using a closed-form Student's t mixture prior for robust unsupervised image prototype discovery.

This paper models low-frequency sound propagation over surrounding topography to evaluate noise pollution caused by AI data centers.

This paper establishes regret bounds and develops adaptive algorithms for prediction with expert advice under heavy-tailed losses.

This paper presents a framework to minimize prediction-set volume in multivariate conformal prediction by learning a Mahalanobis ground cost.

Uses language model agents and deterministic translation to convert TLA+ formal specifications into machine-checkable Verus proofs.

Uses Gröbner bases to simplify pairwise constraints and improve assignment efficiency in constrained k-means clustering.

Introduces bidirectional interleaved pipeline parallelism to improve throughput and activation memory scaling in long-sequence model training.

Applies classifier-free guidance to diffusion policies to separate reinforcement learning exploration from core task capabilities.

Proposes a single-step generative policy for offline reinforcement learning by modeling average velocity instead of iterative denoising.

Formulates a phase-type Hawkes process model with certified convex optimization to measure and predict financial event streams.

Predicts magnetic core losses across materials using partial optimal transport to align operating conditions with limited data.

Identifies how non-separable priors in prior-data fitted networks affect predictive accuracy under model misspecification.

Steers frozen diffusion models toward black-box materials design goals using evolutionary search along sampling trajectories.

Introduces an algebraic framework proving information loss asymmetry when decomposing and reconstructing joint probability distributions.

Develops an episodic sequential decision agent providing hard budget guarantees and sublinear regret bounds using conformal prediction.

Shows the PATTERN graph benchmark can be solved by memorization because pattern subgraphs are reused across splits.

Provides statistical error bounds for dense and sparse mixture-of-experts models, showing sparsity reduces compute rather than sample complexity.

Formulates an online sampling algorithm that adjusts temperature dynamically under cross-entropy budgets to prevent repetitive language model output.

Derives a verifier-tilted temperature sampling formula that optimizes pass-at-k metrics under entropy diversity and plausibility constraints.

Reconstructs magnetic reconnection plasma states from sparse multimodal observations using a residual 3D U-Net architecture.

Demonstrates theoretical sampling failures and kinetic traps when training unconstrained machine-learned interatomic potentials from energy and force data.

Enables prefill-free multi-agent communication across different language model families using cross-model key-value cache transfer.

Uses reinforcement learning to place token anchors that guide text generation in frozen diffusion language models.

Proposes a sequential selection algorithm to prune redundant random features for multiclass classification with theoretical guarantees.

Introduces a recurrent latent graph transformer that solves capacitated vehicle routing problems by updating edge representations iteratively.

Presents an energy-regularized regression method to learn heat equation solutions from noisy physical data and proves estimation error bounds.

Analyzes how data and network structure alignment enables personalized model groups in decentralized learning without explicit personalization algorithms.

This paper investigates error compounding and scale-state feedback mechanisms during recurrent-state quantization in linear-attention and state-space language models.

This paper proposes a graph differential equation and transformer architecture to perform joint data fusion and state estimation in power distribution grids.

This paper introduces a graph-guided layer fusion method that identifies and merges functionally redundant non-adjacent layers to compress language models.

This paper presents an invariant lexicographic canonicalization method to resolve rotation and permutation ambiguities in 3D point cloud representations.

This paper attaches reinforcement learning control heads to a pretrained physics foundation model for closed-loop fluid-flow control.

This paper proposes a game-theoretic reinforcement learning framework to defend virtual power plant resources without requiring grid network topology knowledge.

This paper formulates sequential test-time memory updates as latent variables within a predictive coding framework without backpropagation through time.

This paper studies stability thresholds and convergence boundaries of non-Euclidean gradient descent on quadratic objectives.

This paper proposes a zeroth-order fine-tuning method for language models that optimizes within task-adaptive learnable low-rank subspaces.

This paper derives theoretical bounds on Hoffman constants to analyze the conditioning and convergence of sparse regularized optimization problems.

This paper introduces an observation speculation framework using small world models to accelerate reinforcement learning rollouts for tool-using agents.

This paper introduces a dual-latent variational autoencoder that separates temporal dynamics from cross-sensor correlations for time-series anomaly detection.

This paper introduces an inference-time scaling method using paired denoising and renoising to sample from target distributions despite imperfect verifiers.

This paper derives theoretical conditions on what decision-time history data releases must retain to ensure valid off-policy evaluation.

This paper presents a discrete diffusion model that operates at the fragment level to generate variable-sized molecular graphs.

This paper proposes a few-shot meta-ensembling framework that combines zero-cost proxies to rank neural architectures efficiently.

This paper presents a deep mixed model combining neural networks with trainable subject-specific effects for longitudinal data prediction.

This paper uses a neural operator trained on reverse-time-migration images as a preconditioner to accelerate full-waveform inversion.

This paper introduces learned communication adapters and action heads to coordinate multiple vision-language-action robotic policies during collaborative tasks.

This paper introduces an annotated audio dataset of jazz recordings for beat tracking, chord estimation, and music structure analysis.

This paper introduces a reinforcement learning framework that trains coding agents to anticipate execution outcomes before taking tool actions.

This paper proposes a voice-conversion system using codec language models to preserve speech emotion while anonymizing speaker identity in real-time streams.

This paper analyzes contrastive representation learning theoretically by framing inner-product scores as spectral approximations of pointwise mutual information.

This paper introduces a flow matching framework that decouples backbone generation from sequence and side-chain design for target-conditioned cyclic peptide synthesis.

This paper presents a transform-based framework for structured pruning in language models that optimizes basis selection using activation and weight statistics.

The paper proposes a generation-aware singular value decomposition method to compress large language models while preserving generative output quality.

The paper proposes using vision-language models as success detectors rather than reward regressors for embodied reinforcement learning.

The paper proposes a Riemannian graph tokenizer that quantizes structural patterns to support cross-domain graph knowledge transfer.

The paper develops a neural network constrained by a personalized four-parameter physiological model to forecast blood glucose levels in diabetes.

The paper demonstrates that convolutional neural networks can classify transmitter digital signal processing implementations from over-the-air radio signals.

The paper presents a continuous joint optimization framework for support vector machines that updates multipliers and kernel hyperparameters together.