Addresses preconditioner drift in federated second-order optimizers on non-IID data through curvature alignment techniques.
Proposes AdsorbFlow, energy-conditioned flow matching for fast adsorbate placement prediction on catalytic surfaces.
Bridges GMPO and SAPO by combining sequence-level importance sampling with soft clipping alternatives for improved LLM policy optimization.
Benchmarks graph coarsening trade-offs for GNNs applied to clock tree synthesis in electronic design automation.
Introduces H-GRAMA for merging heterogeneous GNN architectures through routing and message aggregation without retraining.
Proposes Soft Adaptive Policy Optimization (SAPO) replacing hard clipping with smooth sigmoid gate functions to stabilize LLM training and reasoning in GRPO framework.
Framework combining active perception and disentangled representations for continual and few-shot learning without destructive interference.
Method for training LLMs to reason using off-policy reinforcement learning, addressing policy lag in distributed training architectures.
Analysis showing isotropic Gaussian representations improve stability in deep RL under non-stationary training dynamics.
Spiking graph neural network with predictive coding framework for out-of-distribution generalization in dynamic web environments.
Latent ensemble variational data assimilation method for long-range geophysical forecasts using differentiable dynamics.
Federated learning approach for causal representation learning in state-space systems enabling decentralized counterfactual reasoning across networked assets.
Optimal impulse control framework for concentrated liquidity provision in decentralized exchanges using Stein thresholds.
Physics-informed system for analyzing conformational state transitions in Aβ₄₂ protein trajectories for Alzheimer's disease research.
Benchmark combining general reasoning LLMs with domain-specific time-series knowledge for improved time-series diagnostic reasoning tasks.
Evaluation of conformal prediction methods for EEG classification handling distribution shifts in healthcare without standard i.i.d. assumptions.
Interactive browser-based educational platform for learning Federated Learning concepts with real-time visualization of heterogeneous data effects.
Bilevel optimization analysis showing convergence benefits of fewer domain weight updates over longer training horizons.
Variational framework for diffusion models with matrix-valued anisotropic noise schedules that jointly train score networks.
Method for modeling irregular multivariate time series with missing values using time-agnostic summary statistics instead of deep learning.
Investigation of grokking phenomenon in neural networks learning multiplication in finite-dimensional algebras beyond group operations.
Theoretical analysis of sample complexity bounds in replicable realizable PAC learning using Cayley graphs and spectral analysis.
Leap+Verify applies speculative execution to accelerate neural network training by predicting and validating future weights across detected regimes.
Investigates adversarial attacks on deep reinforcement learning agents using advantage-based temporal perturbations.
Compares XGBoost, Random Forest, and TabNet for radiation dose estimation in nuclear safety using interpolation-driven ML approaches.
SME-HGT uses heterogeneous graph transformers to predict high-potential small and medium enterprises using public data from SBIR Phase I awardees.
ISO-Bench benchmark evaluates coding agents on real-world LLM inference optimization tasks from vLLM and SGLang frameworks with 54 curated tasks.
Coordinate Ascent Variational Inference algorithm for Bayesian MIDAS regression with bilinear structure.
Framework for evaluating discrete diffusion language models by separating sampler-induced error from denoiser approximation error.
VecFormer improves graph transformers with token-level attention to reduce computational complexity and improve generalization.
Compositional planning with world models enabling agents to compose pre-trained policies for solving complex tasks.
Study of how data anonymization impacts performance of content-based image retrieval systems.
Analysis of quantum kernel methods under data corruption with introduction of Spectral Phase Encoding technique.
Neural architecture for high-resolution spatiotemporal air quality forecasting in Delhi using four years of data.
Study of representational dynamics in minimal continual learning agents across multiple task executions.
PaReGTA encodes temporal electronic health records using LLMs with lightweight fine-tuning for domain adaptation.
PerturbDiff uses functional diffusion models for predicting cellular responses to perturbations in biology.
GRAPHIC: network science approach to visualize and analyze class confusion patterns in deep learning models.
Representation learning approach to address confounding in Mendelian Randomization epidemiological research.
Unsupervised anomaly detection for network intrusion detection using β-VAE on NSL-KDD dataset.
Bayesian meta-learning method using causal embeddings to improve adaptation to out-of-distribution tasks.
Position paper advocating adoption of data frugal machine learning practices to reduce computational and environmental costs.
Method to recover the correct ordering of shuffled neural network layers using only the dataset and layer information.
Novel Hessian estimators using random direction stochastic approximation for optimization with noisy measurements.
DSDR: dual-scale diversity regularization method to improve exploration in LLM reasoning tasks with reinforcement learning from verifiers.
MSFlow: generative model using flow matching to perform de novo molecular structure elucidation from mass spectrometry data.
Deep learning framework for predicting microstructure evolution in materials science as alternative to expensive phase-field simulations.
Proposes uncertainty-aware rank-one MIMO Q-network for offline reinforcement learning to address extrapolation error from out-of-distribution data.
Analysis of LoRA parameter-efficient fine-tuning in differentially private federated learning for large vision and language models addressing privacy-utility trade-off.
Extends diffusion model role in robust classifier training beyond synthetic data generation to leverage internal feature representations for adversarial robustness.