GRAIL uses retrieval-augmented inference with hyperbolic geometry to improve LLM predictions of clinical events from patient trajectories.
FLAC proposes a likelihood-free reinforcement learning framework using kinetic energy regularization for diffusion and flow matching policies.
TRACE applies agentic context evolution to LLMs for temporal reasoning over streaming electronic health records, improving clinical prediction without fine-tuning.
Amortized Reasoning Tree Search decouples proposal and decision-making in LLMs for enhanced reasoning without suppressing valid paths.
Contrastive learning approach for forecasting aircraft wake vortex trajectories from sparse LiDAR measurements.
Theseus: training-free method for transferring task-specific model updates across different neural network architectures.
Category-level causal feature selection method for multi-label classification using fine-grained causal mechanisms.
Variation Calibration Error metric extending Expected Calibration Error for assessing classifier confidence calibration.
Variational autoencoder ensemble method for anomaly detection in streaming data with concept drift adaptation.
MAUNet neural architecture for bias correction and downscaling of satellite and climate model precipitation data.
Federated Granger Causality framework for causal inference from distributed time-series data with uncertainty quantification.
Machine learning classification of meditation states using fMRI brain imaging data and regional homogeneity analysis.
Wind power forecasting using gradient boosting trees and weather ensemble forecasts for probabilistic day-ahead predictions.
Symbolic regression method incorporating scientific priors to prevent pseudo-equations and ensure consistency with physical principles.
Hierarchical RL approach for dynamically learning temperature sampling policies in LLMs from internal states and task rewards.
Adversarial training method for robust constrained reinforcement learning against temporally coupled perturbations in safety-critical domains.
Detection system for distinguishing AI-generated from human-authored text to prevent misinformation and content fraud.
Geometric approach to imbalanced classification addressing topological intrusion of majority class into minority manifold.
Quantization-aware collaborative inference framework for deploying large embodied AI models on resource-limited agents.
Detection method for identifying when flow matching models extrapolate beyond training data in conditional generation tasks.
Memory-efficient structured backpropagation for on-device LLM fine-tuning balancing gradient accuracy and memory constraints on mobile.
Memory-efficient fine-tuning technique for LLMs on mobile devices via layer-cyclic selective backpropagation, reducing gradient computation.
Pre-training framework using domain-specific expert encoding for unified modeling of homogeneous and heterogeneous graphs.
Solution to diversity collapse in self-play LLM training where challenger-solver loops degrade over iterations despite initial gains.
Study on which discrete algorithms graph neural networks can learn, advancing understanding of neural algorithmic reasoning capabilities.
Method for preserving LLM unlearning effectiveness through low-rank adaptation when models undergo post-training quantization to 4-bit precision.
Parameter-efficient fine-tuning framework for classifying humanitarian disaster information from social media using lightweight LLMs in resource-constrained settings.
RaSD framework for pre-training medical image foundation models entirely on synthetic data using randomized generation.
Actor-critic algorithm for risk-averse multi-agent reinforcement learning in general-sum Markov games with convergence guarantees.
Gradient descent acceleration for quantum Lyapunov Control in QAOA to reduce training overhead and mitigate barren plateaus.
Reproducibility study of DragDiffusion, a diffusion-based method for interactive point-based image editing with spatial control.
ARMOR: Vision language model system for robotic failure detection using adaptive multi-task learning without extensive human annotations.
MiDAS: Open-source platform-agnostic system for time-synchronized multimodal data acquisition in robot-assisted minimally invasive surgery.
CC-Delta: Sparse autoencoder defense against LLM jailbreaks using context-conditioned steering of safety-relevant features.
CacheMind: Conversational tool using RAG and LLMs for semantic reasoning about CPU cache replacement and trace analysis.
RBCorr method to correct response biases in language models tested on 12 open-weight models to improve accuracy.
Gaussian Process approach with gradient information for real-time quadrotor dynamics modeling using state-space partitioning.
Toolkit for scaling multi-vector visual retrieval using training-free pooling and multi-stage search; practical RAG implementation.
AI agents combining LLMs with operations research for inventory control; demonstrates human-LLM-OR complementarity in decision-making.
Evaluates HiFloat low-bit formats for LLM inference on Ascend NPUs; compares INT8 and 4-bit quantization strategies.
Theoretical analysis of regularization-sharpness tradeoff for overparameterized linear interpolators, extending bias-variance concepts.
Open-source vision-language-action model for robotics with real-time execution, trained on cross-embodiment trajectories and vision-language data.
Probing method for vision encoders on multi-channel imaging data with varying channel configurations.
Annotation-free painting restoration framework using synthetic craquelure generation with Bézier curves for unsupervised learning.
BERT-based model with mixture-of-experts routing for aspect-based sentiment analysis in Persian tourism reviews.
RAT-Bench benchmark evaluates text anonymization tools used before LLM training, assessing effectiveness at preventing re-identification.
Framework mapping neural attention computations to programmable network dataplanes with symbolic constraints for trustworthy inference.
Evaluates robustness of object detection models in autonomous vehicles under adverse weather using synthetic data augmentation.
Method for improving reasoning in multimodal LLMs by addressing unreliability in chain-of-thought with interleaved image-text reasoning.
Causal framework for debiasing click-through rate prediction in online advertising when marketing interventions like coupons introduce confounding bias.