2787 papers
This paper proposes using language models, particularly BERT, to capture the semantic meaning of phrases in malicious URLs, achieving a significant improvement in multiclass classification accuracy up to 99.7% compared to traditional methods that rely solely on lexical and statistical features.
This study presents a formally grounded, closed-loop framework that leverages Design Science Research and style-specific AHP-derived weighting to overcome the limitations of uniform evaluation in AI-assisted martial arts training, thereby enabling real-time pose assessment and significantly reducing deviations through knowledge-driven corrective feedback across five distinct Chinese martial arts styles.
AURA is an edge-assisted ZK-Rollup framework that enables privacy-preserving IoT transactions by offloading heavy cryptographic computations to edge nodes and utilizing a novel binary-tree aggregation protocol to compress 64 device-level proofs into a single, low-cost, and energy-efficient on-chain verification.
This study demonstrates that despite extensive experimentation with diverse preprocessing strategies, classical machine learning models, domain adaptation techniques, and contrastive deep learning frameworks, Gestational Diabetes Mellitus (GDM) prediction models trained on single-population cohorts universally fail to generalize across clinical settings due to severe concept drift in the conditional relationship , with performance only recoverable when training data includes both source and target populations.
This paper demonstrates that calibration-conditioned noise models reveal depth-1 QAOA to be universally inferior to classical greedy heuristics across all tested Max-Cut instances and IBM device generations, while establishing that device suitability is predictable from graph topology but indistinguishable between individual healthy chips of the same generation, thereby refining the QSE framework to prioritize generation-level hardware selection over per-device optimization.
This paper proposes MetaFiLM-HTR, a writer-adaptive meta-learning framework that combines FOMAML with FiLM conditioning and explicit gradient routing to enable rapid test-time adaptation for historical and multilingual handwritten text recognition, achieving significant character error rate reductions across diverse datasets through optimized architecture and self-supervised auxiliary tasks.
The paper introduces r2py, a human-supervised, AI-agent framework that orchestrates a multi-phase translation and rigorous verification process to convert R statistical packages into native Python libraries while preserving compiled code and ensuring numerical correctness against the original implementations.
This paper introduces a counterfactual benchmark demonstrating that while ambiguous tool timeouts impose a 50% success ceiling on LLM agent recovery, implementing stable idempotency contracts enables perfect recovery, whereas relying solely on status information yields only partial improvement.
This paper introduces TIERBLEED, a harm-free measurement protocol demonstrating that RL-aligned safety policies in the Gemini family exhibit significant intra-family drift across six reasoning tiers, where framing techniques and multi-tier/seed strategies can exploit these inconsistencies to reproduce mild insults with success rates far exceeding direct refusal rates.
This paper proposes an efficient, attention-gated hybrid Transformer-CNN framework that integrates a hierarchical Mix Transformer encoder with ASPP and a custom Cross-Scale Multimodal Attention Gate to achieve state-of-the-art plant disease segmentation performance on the challenging PlantSeg dataset while maintaining a compact model size suitable for resource-constrained edge devices.