Study of spectral ghost phenomenon in self-supervised learning. Analyzes representation learning using unlabeled data for downstream task transfer.
RL-based preference optimization for generative recommenders. Proposes SAGE to address symmetric conservatism failure in list-wise ranking with multi-objective feedback.
Gauss-Newton natural gradient descent method for shape learning addressing ill-conditioning in implicit neural surfaces.
Pareto-Conditioned Diffusion framework for offline multi-objective optimization using conditional sampling.
Deep learning models for electricity price forecasting in volatile markets with extreme price spikes.
Analysis of optimization instability in deep networks caused by singularities in parameter and representation space.
Bayesian approach for adaptively determining neighborhood aggregation scope in graph neural networks.
Method for erasing harmful concepts from diffusion models while maintaining robustness and generation quality.
tLoRA framework for efficient multi-LoRA training on frozen LLM backbones with elastic shared super-models.
Learnable Chernoff Baselines for efficient inference-time alignment of generative models using reward guidance.
Flow models for efficient generative modeling on Riemannian manifolds with reduced inference evaluations.
Thermodynamic framework analyzing transformer attention through Lagrangian mechanics and entropy minimization.
LLaDA2.1 text diffusion improvement combining token-to-token and mask-to-token editing for faster generation.
Formalization of attention-constrained inference for screening and verifying candidates under limited review capacity.
Analysis showing transformer learning dynamics collapse onto low-dimensional manifolds despite high parameter dimensionality.
Multilingual embedding models using contrastive learning on diffusion-pretrained backbone for web-scale retrieval.
HiFloat4 block floating-point format for efficient LLM inference, achieving 4.5 bits per value with three-level scaling.
Statistical method for estimating consumer preferences when purchase observations are incomplete.
Deep neural network approach for automatic linear graph layout optimization and visualization.
Online tensor inference method for real-time processing of sequentially arriving high-dimensional data with statistical capabilities.
ML research on dynamic assortment optimization with dual contextual information for e-commerce recommendation systems.
ML research on continuous-time Q-learning for jump-diffusion models using Tsallis entropy regularization instead of Shannon entropy.
Representation learning approach for design-based weighting in causal inference without outcome information.
Minmax trend filtering generalizes total variation denoising through local minmax/maxmin formulas for improved signal smoothing.
SaVe-TAG uses LLM-based interpolation to address long-tailed class imbalance in text-attributed graphs for GNN generalization.
Theoretical analysis of Transformers as measure-to-measure maps implemented as interacting particle systems on the unit sphere.
Post-hoc method adding probabilistic uncertainty to vision-language models like CLIP to better handle domain shifts in downstream tasks.
Hierarchical retrieval method offering interpretability and efficiency improvements over embedding-based similarity search for large-scale systems.
Variational Bayesian approach for phylogenetic tree inference using coalescent-based variational families.
Active sampling strategy for MRI point-of-care applications using sequential decision making from k-space with fewer measurements.
PoliCon benchmark evaluates LLMs on achieving political consensus objectives using deliberation records from European Parliament.
Sample-specific noise injection strategy for diffusion-based adversarial purification to improve robustness against adversarial attacks.
Statistical diagnosis and training methods for LLM-based web agents addressing multi-step interactions and reducing post-training compute costs.
Analysis of coordinated online behavior using multimodal approaches to detect disinformation campaigns and collective manipulation.
Technical analysis of load-balancing designs for AI training workloads, comparing approaches and establishing optimality bounds for distributed training.
Highlight & Summarize method for RAG systems to prevent jailbreaking and model hijacking of LLMs through prompt injection defense.
ToolACE-MT enables non-autoregressive generation for multi-turn LLM agent interactions with complex function calls, reducing data generation costs.
Standardized multi-layer tissue maps as metadata format for whole slide image AI algorithm development in digital pathology.
Variational quantum approach for solving conic optimization problems defined over sparse graphs in physics and engineering applications.
Bayesian method for decentralized multi-agent reinforcement learning over networked graphs with dynamic neighborhoods and constrained communication.
HEART framework uses emotional cues during test-time scaling to improve LLM problem-solving by preventing repetitive thought patterns through alternating critical and encouraging tones.
Research on physical embodiment constraints for AI agents in eldercare and disaster response scenarios, addressing generalization and care provision under uncertainty.
VoiceAgentBench evaluates speech language models on agentic multi-turn tasks and adversarial robustness beyond isolated capabilities.
Benchmarks six ML models (CNN, LSTM, XGBoost, transformers) for radio link failure prediction in 5G railway networks.
Reinforcement learning framework for beam positioning in LEO satellite constellations using weighted least squares estimation.
Knowledge graph completion using attention mechanisms and diffusion models for few-shot learning on long-tailed relations.
Fine-tuned LLM for depression screening in Nigerian Pidgin English addresses language accessibility in mental health diagnostics.
Feature selection methodology for biomarker discovery in high-dimensional biomedical data with low sample sizes.
Standardized framework for tuberculosis detection from cough audio combining machine learning models with clinical variables and uncertainty quantification.
Framework and tools for building and standardizing provenance-based intrusion detection systems with consistent evaluation metrics.