2787 papers
The paper introduces ScintiGround-38K, a large-scale, provenance-preserving benchmark for grounded visual question answering on paired anterior-posterior bone scintigraphy images, which reveals that while models achieve high raw answer accuracy, they struggle significantly with minority-class detection and strict evidence-based localization.
This paper introduces a lightweight, character-level TF-IDF and Logistic Regression baseline for token-level Hinglish Language Identification that achieves 95.92% accuracy on the COMI-LINGUA dataset, outperforming multilingual models by approximately 7% while providing critical error analysis and a roadmap for future inclusive language technologies in India.
This paper demonstrates that Exclusive Self-Attention (XSA), a mechanism designed to reduce self-referential behavior, remains stable under Muon optimization and yields modest generalization improvements in language models, with performance gains becoming more consistent as model scale increases from 12 to 24 layers.
This paper proposes a parameter-efficient, language-agnostic framework for adapting Whisper models to low-resource domain-specific speech by integrating LoRA-based attention fine-tuning, on-the-fly SpecAugment, and a two-stage decoding strategy with language model rescoring.
Sentinel-Mesh is a neuro-symbolic framework that combines LLM-driven patch synthesis with SMT-based formal verification to autonomously remediate cloud misconfigurations while guaranteeing the elimination of hallucinated patches and ensuring semantic security within a defined abstraction of network exposure, encryption, and data sensitivity.
This paper introduces a unified adaptive framework for image enhancement that generalizes integer-order operators like Sobel and Laplacian into a continuous Grünwald-Letnikov fractional-order spectrum, featuring multi-directional fusion, complex-order phase selectivity, and an AlphaNet-driven adaptive order mechanism to achieve superior texture preservation and noise control across diverse domains.
This paper proposes a conceptual framework for quantifying cybersecurity risk in AI systems by introducing five critical dimensions—Attack Surface, Model Exploitability, Data Integrity, Regulatory Exposure, and Control Maturity—to address the limitations of existing IT-focused tools and establish a shared vocabulary for translating AI-specific threats into organizational decisions.
This paper proposes a probabilistic 2D human pose estimation framework that integrates a Multi-Scale Context Fusion Vision Transformer backbone, an Adaptive Probability Map for dynamic uncertainty modeling, and an Adaptive OKSLoss to significantly improve performance in challenging scenarios involving occlusion and out-of-bound keypoints.
The paper proposes HRMN, a novel time-series learning framework that mimics the primate cortex's hierarchical processing by decomposing signals into intrinsic modes via EEMD, routing transient and persistent components to specialized TCN and Transformer branches based on their Hurst exponents, and achieving state-of-the-art performance across classification, regression, and anomaly detection tasks with significantly fewer parameters than existing models.
This paper proposes CAFAD, a context-aware method that leverages cross-perception attention between flight phases and localized airspace-specific context sets to achieve fine-grained, phase-level anomaly detection for improved flight trajectory safety.