Human-centered architecture for integrating LLM-based cognitive assistants into manufacturing quality management systems.
Interpretable ML framework for predicting non-small cell lung cancer drug response using patient genetic data.
Security research on sentiment steering attacks targeting RAG-enabled large language models and LLM robustness.
Machine learning pipelines for radio astronomy data processing with explainability focus on automating configuration.
YOLO-based deep learning for automated wasp identification with explainable AI integration for taxonomic classification.
Plaza6G platform for experimental trials in 5G/6G networks with AI-assisted orchestration of cloud and wireless resources.
Remote sensing monocular depth estimation using Vision Transformers and diffusion models for real-time processing.
DynamicGate MLP conditional computation framework using learned structural dropout and input-dependent gating for efficiency.
FederatedFactory zero-dependency framework for federated learning in non-IID scenarios using generative one-shot learning.
Age prediction models analyzed for out-of-distribution generalization, bias mitigation, and interpretability with causal implications.
Physics-guided diffusion framework for full-waveform inversion combining score-based generative models with wave-equation simulations.
Fanar 2.0 Arabic generative AI platform built on 256 H100 GPUs at QCRI with sovereign infrastructure and data pipelines.
Study identifying flaws in LLM benchmarks for Icelandic, highlighting issues with synthetic and machine-translated evaluation data.
PlotTwist creative plot generation framework using small language models with specialized training for narrative coherence.
Method for adding persistent memory to frozen encoder-decoder LLMs via trainable adapters in continuous latent space.
IndexRAG approach for multi-hop question answering that performs cross-document reasoning at indexing time using bridge entities.
SF-Mamba state space model for vision addressing non-causal patch interactions with improved computational efficiency.
SlideFormer system for fine-tuning large language models on single GPU via asynchronous engine and sliding window approach.
LenghuSky-8 eight-year all-sky cloud dataset with star-aware masks for astronomical observatories and nowcasting.
EngGPT2-16B open Italian LLM trained on 2.5T tokens, efficient inference with performance comparable to larger models.
CD-FKD cross-domain feature knowledge distillation for single-domain generalization in object detection.
Multi-agent reinforcement learning approach for managing delayed channel state information in multi-satellite communication systems.
DST-Net dual-stream transformer for low-light image enhancement using spatial convolution and feature guidance.
Unlearning method for one-step generative models using unbalanced optimal transport for safer image generation.
LenghuSky-8 millisecond-resolution network dataset for time series foundation models with high-frequency data.
FEAT foundation model with linear complexity for structured data in healthcare, finance, and e-commerce with improved scalability.
DanceHA multi-agent framework for document-level aspect-based sentiment analysis, extracting ACOSI tuples from documents.
CompDiff uses hierarchical compositional diffusion to generate fair medical images across demographic groups and intersections.
EmoLLM framework integrates appraisal-grounded cognitive-emotional reasoning into LLMs for contextually appropriate responses.
Deep learning inverse design for Doherty power amplifiers using CNN surrogate models and genetic algorithms.
Analysis of human-LLM chat logs characterizing delusional spirals and negative psychological effects from extended chatbot interactions.
Manifold-Matching Autoencoders regularize autoencoders by aligning pairwise distances between latent and input spaces.
Research on classifying malicious AI agent skills using repository context to improve detection in skill marketplaces.
REFORGE reveals vulnerabilities in image generation model unlearning through multi-modal adversarial attacks in black-box settings.
BATQuant proposes outlier-resilient MXFP4 quantization via learnable block-wise optimization for deploying MLLMs and LLMs on accelerators.
Frequency-spatial fusion framework for cattle mounting pose estimation in cluttered, occluded environments.
Data-driven perimeter control for urban traffic congestion using machine learning instead of explicit modeling.
Analysis of multimodal LLM-generated natural language explanations for face verification on unconstrained face images.
Omanic, a benchmark for step-wise evaluation of multi-hop reasoning in LLMs with step-level annotations for diagnosing failures.
Investigation of linguistically related language guidance for LLM translation in low-resource settings without large parallel data.
Study of emergent AI agent communities on platforms, analyzing 167k+ agents learning from each other without researcher intervention.
Kestrel, a training-free method for mitigating hallucinations in large vision-language models using grounding and self-refinement.
World action models for embodied control that eliminate test-time future imagination while maintaining action performance.
Resource-aware LLM-based agent reasoning for embodied robots using reinforcement learning to balance computation and action execution.
Computational cost analysis for matrix inversion updates in online outlier detection systems.
Federated learning models for predicting postoperative complications using multi-center healthcare data.
In-context learning improvement for vision-language models using retrieved counterfactuals for better visual reasoning.
SpecMoE mixture-of-experts foundation model for cross-species EEG decoding with spectral-temporal fusion.
Formal model for selecting statements that find common ground across diverse preferences using generative AI.
TurnWiseEval benchmark and analysis of multi-turn vs single-turn LLM capabilities with step-level evaluation.