All papers

The paper presents an invertible normalizing flow method for two-dimensional topographic manifold learning and data visualization.

The paper integrates matrix sketching via frequent directions into adaptive optimization methods to reduce preconditioning computation overhead.

The paper theoretically analyzes how one-layer transformers learn in-context nearest-neighbor prediction when using chain-of-thought reasoning.

The paper evaluates machine learning models on metadata and acoustic features to predict audio popularity on an online sharing platform.

The paper proposes a dynamic auxiliary search strategy to guide symbolic regression solvers by constructing a high-potential functional subspace.

The paper evaluates neural network architectures to downscale ERA5 atmospheric data to high-resolution satellite precipitation estimates across West Africa.

The paper develops a calibrated barycenter framework to accelerate multiobjective optimization over divergence matching problems.

The paper introduces a center-estimated neural posterior estimator to improve sample-size extrapolation in simulation-based inference.

The paper analyzes when feature standardization improves risk bounds for rare subgroups under finite stochastic gradient training budgets.

The paper proposes a post-training tail correction method for diffusion and flow models using radial-angular extreme value decomposition.

The paper introduces GpFlash, an open library designed to speed up Gaussian process training and prediction while reducing memory footprint.

The paper develops a formal performance theory analyzing skill allocation and workflow design in joint human-machine systems.

The paper presents PivotRL, a system that dynamically shifts GPU resources among rollout, reward modeling, and training during post-training.

The paper demonstrates that small-batch stochastic gradient noise restores sequential feature learning in the Muon optimizer.

The paper analyzes sequence composition strategies during electronic health record foundation model pretraining to improve downstream clinical prediction.

The paper formulates continuous symmetry discovery across function families as an inverse symmetry operator problem using differential characteristics.

The paper introduces FORGE, an on-chip datapath that regenerates sparse mixture-of-experts weights from compact seeds during inference.

The tutorial examines machine learning engineering agents from a data management view, highlighting challenges in pipeline execution, data discovery, and debugging.

The paper presents a temporal reinforcement learning framework that selects physiologically aligned time windows for motor imagery brain-computer interfaces.

The paper provides theoretical convergence and tracking error bounds for Matrix Shampoo optimization in nonstationary stochastic settings.

The paper provides two modifications to make tensor decoupling methods more robust to magnitude mismatches and eliminate manual spline tuning.

The paper presents AutoConvexRelax, which learns a reinforcement learning policy to automatically construct valid convex relaxations for nonconvex optimization problems.

The paper proposes a multi-agent diffusion transformer framework to generate realistic air traffic scenes and trajectory patterns from recorded data.

The paper introduces HarnessSR to jointly optimize agent workflows and prompts for discovering symbolic regression equations from data.

The paper presents GenTuRBO, using pretrained generative models as adaptive proposal trust regions for discrete Bayesian optimization tasks.

The paper proves constant-depth rational-weight graph neural networks match the 1-WL expressiveness bound under Schanuel's conjecture from number theory.

The paper proposes a method to learn the intrinsic geometry and diffusion operators of underlying processes from distorted, indirect observations.

The paper introduces a physics-coupled graph neural operator to predict mid-plane stress tensor fields in concrete during projectile penetration.

This paper develops a deferred-decision trajectory optimization controller with conformal prediction guarantees for multi-agent target interception.

This paper introduces an asynchronous population of generative flow networks coupled through soft clustering to improve mode coverage.

This paper uses neural network variational Monte Carlo simulations to demonstrate the formation of honeycomb electron-hole crystals in bilayers.

This paper introduces a model-free variable importance measure that uses active variable selection to prevent nonactive variables from receiving spurious credit.

This paper analyzes in-context classification in linear attention models, establishing theoretical asymptotic bounds separating calibration from noise effects.

This paper introduces an interface for extracting context-rich language model embeddings to improve table foundation models on mixed tabular data.

This paper proposes a nearest-neighbor functional curve completion method for intraday probabilistic revisions of renewable energy forecasts.

This paper presents an integer-only local learning system on edge FPGAs for energy-efficient convolutional neural network training.

This paper presents a masked diffusion model that synthesizes complete household travel survey diaries from tokenized sequences.

This paper provides a formal Lean proof establishing that guarded masked attention pooling is invariant to padding and row permutation.

This paper proposes a method that uses conformal score gradients to guide Markov chain Monte Carlo exploration for fast, calibrated Bayesian inference.

This paper develops and evaluates implicit and explicit reparameterized gradient estimators for Folded Normal and Rice distributions in variational inference.

This paper presents an end-to-end spiking neural network that decodes and reconstructs dynamic visual stimuli directly from biological retinal spike trains.

This paper develops a generalized McDiarmid concentration inequality capturing sample reuse and unbounded differences for stochastic optimization and generative modeling.

This paper introduces an uncertainty-aware property prediction workflow to guide out-of-distribution candidate selection for ionic liquid solvent discovery.

This paper presents a label-free decision head for frozen language models that preserves multiple-choice accuracy with reduced layers and memory.

This paper presents a cleaned and enriched residential floor plan dataset with graph, geometry, and image representations for machine learning.

This paper introduces a coordinate encoder trained on open map data to produce multi-scale location embeddings without requiring imagery.

This paper proposes density-matrix reinforcement learning, representing policies as density matrices to capture relationships and coherences between actions.

This paper introduces a reinforcement learning method that calibrates test pass-rate rewards to create adaptive training curricula for software engineering agents.

This paper introduces an implicit solvation potential providing solvent-dependent corrections to gas-phase machine learning interatomic potentials without retraining.

This paper presents a discrete flow-matching generator that evaluates molecular identity probabilities on demand to accelerate high-fidelity molecular generation.

This paper presents a distillation method using frozen language models to guide molecular flow generators toward natural language discovery goals.

The authors propose an action quantization method that preserves value functions by minimizing reachability error for search-based robot policy learning.

The authors compare federated learning algorithms for hospital mortality prediction and report that evaluation protocol choices outweigh aggregation algorithm differences.

The authors propose a score-based diffusion framework that controls generation speed across different graph frequency bands via colored noise priors.

The authors demonstrate that a single autoregressive neuron using chain-of-thought generation can simulate any feedforward circuit of the same activation type.

The authors provide a theoretical criterion to decide whether neural network layers should receive sign-based or spectral optimization updates during training.

The authors introduce a reinforcement learning method that optimizes language model explanations specifically to improve human downstream decision accuracy.

The authors develop a geometric construction method to build two-hidden-layer neural networks that achieve exact classification on given training sets.

The authors propose a deep unrolling algorithm combining alternating projections with a learned denoiser for phase retrieval under changing measurement conditions.

The authors derive an online-to-offline reduction achieving near-optimal regret bounds for strongly monotone stochastic optimization under semi-bandit feedback.