Template-free retrosynthesis framework using discrete flow matching with structure-aware generation for chemical reaction prediction.
Optimization techniques for graph neural network molecular dynamics simulation on GPUs, improving memory access patterns.
Method for preserving LLM unlearning effectiveness through low-rank adaptation when models undergo post-training quantization to 4-bit precision.
Graph neural networks applied to facility location combinatorial optimization problems with provable performance guarantees.
Neural network approach for recovering unknown functions from data in partial differential equations.
Parameter-efficient fine-tuning framework for classifying humanitarian disaster information from social media using lightweight LLMs in resource-constrained settings.
Ground fault localization for DC-side faults in three-phase TN-earthed photovoltaic systems using simulation-based analysis.
Gradient boosted mixed-model ML framework analyzing AIS data to predict vessel speed in Arctic using two-stage approach.
MusicRecoIntent corpus of 2,291 annotated Reddit music requests with intent labels for music recommendation systems.
Visible and hyperspectral imaging analysis for milk quality assessment as alternative to conventional chemical analysis.
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.
Non-coherent over-the-air computation framework for decentralized optimization in wireless systems without channel state information.
RBCorr method to correct response biases in language models tested on 12 open-weight models to improve accuracy.
Signal quality indices for medical time series like ECGs and PPGs to identify unreliable algorithm outputs in noisy environments.
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.
Theoretical analysis of linear regression with unknown truncation for non-Gaussian features.
NLP methods for discovering semantic structures in psychological questionnaires without response data.
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.
U-Net architecture with additive skip connections for joint image denoising and classification tasks.
Theoretical analysis of regularization-sharpness tradeoff for overparameterized linear interpolators, extending bias-variance concepts.
Deep learning for binary code similarity detection; evaluates robustness against semantics-preserving transformations in cybersecurity reverse engineering.
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.
Rate-distortion analysis for task-oriented source coding, optimizing visual data compression jointly with task model performance.
Multi-objective bandit algorithm showing that multiple good arms across objectives can provide implicit exploration benefits.
Evaluates robustness of object detection models in autonomous vehicles under adverse weather using synthetic data augmentation.
Repeated bilateral trade problem where agent valuations depend on context vectors, requiring learning of unknown preference vectors.
Online learning algorithm for bilateral trade with contextual information about seller and buyer valuations in nonparametric setting.
Method for improving reasoning in multimodal LLMs by addressing unreliability in chain-of-thought with interleaved image-text reasoning.
Theoretical analysis of annealing strategies in variational inference to mitigate mode collapse when learning Gaussian mixtures.
Training-free conditional sampling method for flow matching models using importance sampling and sequential Monte Carlo resampling.
Multi-label Arabic dialect identification using curriculum learning and pseudo-labeling to overcome single-label dataset limitations.
Causal framework for debiasing click-through rate prediction in online advertising when marketing interventions like coupons introduce confounding bias.
Framework analyzing bio-inspired RNN models with chemical synapses and synaptic activation to improve interpretability of recurrent neural networks.
FedHENet extends federated learning to image classification using fixed feature extractors, reducing privacy risks and computational costs in heterogeneous environments.
Curriculum-DPO++ applies curriculum learning to direct preference optimization for improved text-to-image generation model training.
Design-based statistical formulation of random forests with exact variance decomposition for finite-sample analysis.