Technology
A Methodological Framework for Explicit Control of the Speed-Accuracy Trade-off in Brain-Computer Interfaces
This study introduces a classifier-independent framework using Gain-Cons Balance to explicitly control the speed-accuracy trade-off in BCIs. Tuning parameter $\alpha$ enables customizable performance without altering underlying classifiers.
Project SPARROW and the Future of Conservation Technology
Project SPARROW is an open-source, solar-powered biodiversity surveillance system using edge AI and satellite comms. Field tests across four countries demonstrated its ability to autonomously monitor ecosystems for 190 days, enabling scalable, remote ecological tracking.
Enhancing BiGRU with a KAN Block for Legal Document Classification and Summarization
This study integrates Kolmogorov-Arnold Networks into BiGRU models for legal document classification and summarization in low-resource multilingual settings. Results show significant performance improvements, particularly in classification accuracy, over traditional baselines.
V2I Work Zone Geometry Reconstruction with Pose-Conditioned UWB Range Denoising
This study introduces a pose-conditioned UWB denoiser for V2I work zone reconstruction, significantly improving range accuracy and geometry mapping. It achieves a 66.9% MSE reduction, demonstrating robustness against NLOS errors and anchor dropout.
SpikeWFM: Spiking-Aided Wireless Foundation Model for Robust Channel Prediction
SpikeWFM merges spiking neural networks with transformers to enhance wireless foundation models' robustness against noise. It outperforms standard models in convergence speed and channel prediction accuracy.
Versatile Framework with Semantic and Structural guidance for Image Reconstruction from Brain Activity
MindDiffuser reconstructs images from brain activity using semantic and structural guidance. It outperforms existing models across fMRI, EEG, and MEG datasets.
CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations
CardioLens reveals a significant clinical reality gap in MLLMs, showing poor performance on multi-sequence cardiac MRI. Models struggle with complex workflows, often defaulting to common abnormalities rather than accurate diagnosis.
VDSB-GWSyn: Diffusion Schr\"{o}dinger Bridge for Controllable and Anatomically Feasible Guidewire Synthesis in Coronary Angiography
VDSB-GWSyn uses diffusion Schrödinger bridges to generate anatomically plausible guidewires for coronary angiography. This synthetic data significantly improves endpoint localization accuracy, aiding robot-assisted PCI.
A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity
LLMs and human EEG share a common valence axis, yet aligning them degrades performance. This "saturation regularity" shows extra supervision distorts already saturated neural representations.
Automatically Differentiable Nonlinear Tensor Networks (ADNTNs) for Exponential Compression of Deep Neural Networks
ADNTNs achieve exponential DNN compression (up to 77,000x) via optimized core tensors, maintaining or improving accuracy. However, automatic differentiation does not eliminate the inherent computational costs of tensor contractions.
Multimodal Music Recommendation System using LLMs
This study introduces a multimodal LLM-based music recommendation system integrating audio, lyrics, and metadata. It significantly outperforms ID-only baselines, though cross-modal fusion complexities remain.
AI-PROPELLER: Warehouse-Scale Interprocedural Code Layout Optimization with AlphaEvolve
AI-PROPELLER uses AlphaEvolve to optimize warehouse-scale binaries via interprocedural layout. It achieves 0.23-1.6% performance gains on real hardware, marking the first industrial application of this technique.
Foundation-Preserving Adaptation via Generalized Rayleigh-Quotient Optimization
FoLoRA uses generalized Rayleigh-quotient optimization to balance adaptation and preservation, outperforming baselines in maintaining non-target competencies while enhancing downstream task performance.
On Effectiveness and Efficiency of Agentic Tool-calling and RL Training
This study reveals that tool-calling evaluations are highly sensitive to implementation details, while RL training suffers from computational waste. We propose two techniques to accelerate training with significant speedups and maintained performance.
Generative AI and Digital Ecosystem Resilience: A Proactive Lifecycle-Based Survey
This survey proposes a proactive, lifecycle-based framework using the C5 model to detect emerging synthetic threats. It integrates socio-technical and computational methods to enhance digital ecosystem resilience against GenAI-driven adversarial campaigns.
XAI-SOH-FL: Enhancing SOH-FL with Adaptive Aggregation and Explainable AI for Intrusion Detection in Heterogeneous IoT
XAI-SOH-FL enhances IoT intrusion detection via adaptive aggregation and SHAP explainability, achieving 94.12% accuracy. It outperforms SOH-FL with faster convergence and transparent decision-making.
Adaptive data selection improves wearable prediction under low baseline performance
Adaptive data selection significantly boosts wearable prediction accuracy for users with low baseline performance, but offers little benefit to those already performing well.
Geodesics with Unified Tangent-constrained Priors and Curvature Regularization
This unified geodesic framework combines tangent-constrained priors with curvature regularization to prevent shortcuts in complex segmentation. It enhances shape fidelity and robustness against weak boundaries using efficient HJB PDE solvers.
Geometric Erasure by Contrastive Velocity Matching in Rectified Flows
GEM is a concept erasure framework for Rectified Flow models that bridges trajectory-based unlearning and teacher-guided erasure. It uses contrastive velocity matching to suppress unwanted concepts while preserving benign generation.
Completion at the Boundary (CaB): Deployable Switching with Completion-Aware Control under Limited Calibration
Completion at the Boundary (CaB) improves VLA task switching by preserving boundary evidence via phase tokens. This deployable method enhances composite execution and handoff quality under strict calibration constraints.