Benchmark comparing zero-shot text classification across cross-encoders, embedding models, rerankers, and LLMs for matching texts to label descriptions.
arXiv paper on personalized federated learning (PFL) using multi-objective optimization to train customized models across clients with heterogeneous data.
arXiv paper on decentralized orchestration architecture for distributed AI/IoT across heterogeneous resources spanning edge and cloud platforms.
arXiv paper on adaptive graph-enhanced multi-agent reinforcement learning (AGMARL-DKS) for intelligent Kubernetes scheduling balancing stability, utilization, and costs.
arXiv paper on continual learning for vision-language models using semantic-geometry preservation to prevent catastrophic forgetting across tasks.
FlashMotion enables few-step trajectory-controllable video generation using distillation techniques to reduce computational overhead of multi-step denoising.
Proof-Carrying Materials framework provides falsifiable safety certificates for machine-learned interatomic potentials in materials screening applications.
IndexCache accelerates sparse attention in LLM agentic workflows by reusing cross-layer indices, improving inference speed and serving costs for long-context applications.
HiAP proposes hierarchical auto-pruning for Vision Transformers reducing computational demands for edge deployment via multi-granular structured pruning.
Method using fuzzy rules to interpret contrastive text embeddings in domain-specific applications like legal and medical records processing.
BiGain presents token compression framework for diffusion models balancing generation quality and classification via frequency separation, training-free and plug-and-play.
Study examining reasoning LLMs used as judges for evaluating non-verifiable domains in post-training, testing inference-time scaling benefits for policy training.
Spatial-TTT proposes test-time training for streaming visual spatial understanding from video, addressing how spatial information is maintained over unbounded streams.
Graph deep learning model for drug response prediction and biomarker identification using heterogeneous drug-cell-gene networks with attention.
Geometric analysis of ReLU networks using Data Information Matrix to understand data manifold structure and singular foliations.
Gradient-free variant of Stein Variational Gradient Descent combining evolution strategies for sampling from unnormalized distributions.
Orthogonal learner for quantifying aleatoric uncertainty in treatment effect estimation from observational medical data.
Finance-informed neural network for option pricing and hedging using self-supervised replication objective based on dynamic hedging theory.
General Time-series Model with frequency-domain attention for enhanced representation learning on diverse time-series downstream tasks.
Higher-order guided diffusion model for graph generation that captures non-Euclidean topology using higher-order graph structures.
Riemannian Gaussian Variational Flow Matching for generative modeling on manifolds applied to material and protein design.
Aggregation-free federated learning method for medical image classification using multi-dimensional similarity knowledge distillation across heterogeneous client models.
Framework for compressing large LLM-based ReAct agents into smaller student models while preserving reasoning and action consistency.
Framework for generative modeling with enforced physical constraints using split augmented Langevin sampling for scientific applications.
Hierarchical differential model for inferring system degradation from sensor data by disentangling slow and fast temporal dynamics.
Text-trained LLMs perform zero-shot extrapolation of PDE dynamics, revealing three-stage in-context learning mechanism for spatiotemporal forecasting.
Mathematical study of Busemann functions in Wasserstein space with applications to geometric machine learning and data slicing.
Development of conformal prediction method that ensures counterfactual fairness in prediction sets for fair decision-making under uncertainty.
Zeroth-order optimization approach for continual learning that improves memory efficiency and addresses plasticity-stability tradeoffs without gradient computation.
Research on unifying in-context learning and activation steering as instances of a broader framework using belief dynamics to control LLM behavior at inference time.
Introduces RAT+, structured dilated attention architecture that enables sparse inference while maintaining long-range connectivity and accuracy.
Proposes controllable exploration strategy for RLVR training of multi-modal LLMs to address entropy collapse and policy degradation.
Introduces FlashOptim, memory-efficient optimizers for mixed-precision neural network training reducing per-parameter memory requirements.
Shows preference labels in LLM-as-judge training can function as covert communication channels, challenging assumptions about semantic supervision.
Investigates tokenizer pretraining impact on physics foundation models for emulating complex multiphysics phenomena in data-limited settings.
Structure-aware set transformers with temporal and variable-type attention for asynchronous clinical time series in EHR data.
Analyzes how MDP design choices (state composition, rewards, dynamics) affect sim-to-real transfer in reinforcement learning for industrial control.
Instance unlearning method for diffusion models removing specific outputs without text prompts, addressing unpromptable undesired generations.
Proposes iterative selection of Gaussian mixture priors to prevent posterior collapse in variational autoencoders.
AI system analyzing police bodycam footage at scale to assess officer-public interactions and improve government accountability.
Generative Predictive Control method augments frozen diffusion policies with action-conditioned world models for improved robot control without retraining.
Applies multi-agent reinforcement learning to greenhouse gas offset credit markets for emissions control and carbon project trading simulation.
Data-driven survey identifying 14,648 papers on LLM limitations from 2022-2025 using automated classification and expert validation across 250,000 academic papers.
Novel algorithm for multi-agent reinforcement learning using uncertainty quantification and selective exploration to improve sample efficiency in joint action spaces.
Research paper on measuring whether LLMs comprehend user intent beyond surface-level text patterns, addressing training-inference gaps in language models.
Reinforcement fine-tuning approach for LLMs applied to point-of-interest recommendation with improved semantic indexing.
Open benchmark suite comparing paired encoder and decoder architectures for NLP tasks with controlled parameter counts.
Adapter parameters for efficient multi-task LLM inference on-device via task merging for compositional learning.
Agentic Design Review System orchestrates multiple AI agents to collaboratively analyze graphic designs with meta-agent coordination.
arXiv paper analyzes theoretical limitations of embedding-based retrieval for diverse tasks including reasoning and code generation.