Theoretical analysis of single-loop stochastic bilevel optimization convergence for meta-learning and hyperparameter optimization.
FlexGuard proposes continuous risk scoring for LLM content moderation that adapts to varying strictness levels across platforms and time.
FedRot-LoRA addresses rotational misalignment in federated LoRA fine-tuning of LLMs, improving communication-efficient training on decentralized data.
Diffusion-based method for time series anomaly detection using selective denoising instead of conditional reconstruction strategies.
MoST contrastive learning method for disentangled mode-specific representations in multi-mode tensor time series.
Geometric analysis of transformer training trajectories revealing low-dimensional drift direction and transverse oscillatory dynamics.
BDGxRL uses Diffusion Schrödinger Bridge to address dynamics gaps in cross-domain reinforcement learning without target reward supervision.
OPTIAGENT uses LLM-based agentic framework with physics-driven optimization for automated optical design and lens system configuration.
MAGE multi-scale autoregressive generation framework for offline RL addressing long-horizon tasks with sparse rewards via hierarchical decomposition.
Provable identifiability framework for nonlinear multi-view canonical correlation analysis via subspace identification.
Learning-based pathfinding using neural networks to approximate informed heuristics for grid-based search across different map topologies.
MPU framework for privacy-preserving knowledge unlearning in LLMs without sharing server parameters or client forget sets.
Actor-critic pretraining approach for PPO that leverages expert data to reduce environment interactions required for RL training.
Theoretical investigation of offline reinforcement learning with general function approximation and parametric policies beyond state-wise methods.
Q-learning approach for learning safe policies from expert demonstrations with unknown constraints in constrained MDPs.
FedNSAM addresses sharpness-aware minimization in federated learning under high data heterogeneity, ensuring both local and global model flatness.
Ultra-lightweight neural network for automated sleep stage classification from multimodal polysomnography data.
Theory of graph classification under domain shift using random graph models and domain adaptation techniques for structured data.
LK Losses directly optimize acceptance rate in speculative decoding for LLM inference, improving upon KL divergence proxy objectives for draft model training.
Hierarchical Concept Embedding Models improve interpretability of deep neural networks by mapping inputs to human-interpretable concept representations with inter-concept relationships.
Learns optimal generation orders for masked discrete diffusion models via variational inference to balance parallel generation and sample quality.
Proposes Intrinsic Lorentz Neural Network for fully intrinsic hyperbolic geometry operations on hierarchical data representations.
Transfers knowledge from multimodal neuroimaging to speech analysis for early Alzheimer's disease screening via speech-based classifiers.
Outlines vision for foundation world models as persistent compositional representations enabling agents to learn and adapt in open worlds.
Theoretical analysis showing InfoNCE contrastive loss induces Gaussian structure in learned representations for foundation models.
Proposes RewardUQ framework for uncertainty-aware reward models in LLM alignment that reduces annotation costs and prevents overoptimization.
Introduces pathsig, a PyTorch-native GPU-accelerated library for computing path signatures as trainable features for sequential data.
Develops high-dimensional node embedding method using non-linear dimension reduction and random walk co-occurrence for graph tasks.
Proposes ACWI framework that adaptively balances intrinsic and extrinsic rewards online for sparse reward reinforcement learning exploration.
Introduces Neural Diffusion Intensity Models using variational framework with neural SDEs for intractable Cox process inference.
Surveys agentic AI systems with planning, tool use, and self-management capabilities applied to Open RAN network control and optimization.
Studies best arm identification problem with heterogeneous resource costs and constraints across multiple resource types.
Proposes explainable AI method for discrete token inputs like text using attribution highlighting to identify important tokens in transformers.
Develops sandwiching polynomial approximators for learning with distribution shift and contaminated data in low intrinsic dimension settings.
Applies multi-objective reinforcement learning to optimize container consolidation in human-robot collaborative fulfillment centers.
Uses normalizing flows to estimate density ratios between intractable distributions with applications to genomics data analysis.
Proposes federated learning approach for anomaly detection in heterogeneous IoT networks while preserving privacy through distributed training.
Compares classical logistic regression and MLPs with variational quantum classifiers on XOR problem using quantum superposition principles.
Investigates trade-off between regret minimization and statistical power in combinatorial multi-armed bandits using Pareto optimality framework.
arXiv paper benchmarking general-purpose time-series foundation models for zero-shot transportation forecasting across multiple datasets.
arXiv paper on Latent Manifold Compaction for unsupervised harmonization of histopathology images across different batch effects and scanners.
arXiv paper proposing Web-Knowledge-Web pipeline for discovering suppliers in specialized industries via iterative web crawling and knowledge base integration.
Analyzes limitations of standard identifiability metrics (MCC, DCI, R²) on synthetic benchmarks, revealing implicit structural assumptions in representation learning evaluation.
Memory caching architecture enabling RNNs with growing memory capacity and subquadratic complexity as alternative to Transformers for sequence modeling.
Low-rank approximation method (LoRA-Pre) for optimizer memory efficiency in Adam and Muon, reducing overhead for large language model training.
Agentic RL system using LLMs for high-performance CUDA kernel generation at scale, outcompeting traditional compiler-based approaches.
Framework establishing universal approximation properties for shallow and deep neural networks on non-Euclidean topological spaces.
Graph reinforcement learning approach to moderate opinion polarization in social networks under Friedkin-Johnsen model with improved scalability.
Methodology for dynamic neural networks using isotropic activation functions enabling real-time architectural growth and shrinkage via symmetry-principled primitives.
Frequency-aware diffusion model for detecting rare lesions in CT scans with extreme class imbalance using controlled synthetic augmentation.