801 papers
This paper demonstrates that the sharp "emergent" jumps in exact-match accuracy observed in multi-digit integer addition are not genuine discontinuities in model capability but are instead fully predictable artifacts resulting from smooth, power-law improvements in per-token accuracy compounded by the nonlinearity of requiring all digits to be correct.
This paper introduces the cross-expert gradient alignment metric (Γ) to demonstrate that standard Mixture-of-Experts training, particularly in Mixtral 8×7B, is fundamentally driven by a structural zero-sum competition caused by the combined negative effects of softmax coupling and top-k normalization, which creates a persistent negative alignment attractor that can only be overcome by removing these specific architectural constraints.
This paper proposes a novel community detection model that leverages the Skellam Nutcracker Optimization Algorithm (S-NOA) and a SwishSERF Attention Spatial Drop-Macro Unit CNN (S2ASD-MUNet) to effectively analyze inter and intra-layer dependencies, achieving a high modularity score of 0.962.
This paper introduces MESH4ID, a headset-centered virtual reality pipeline built on Unity/OpenXR that enables users to capture, manage, and reconstruct indoor environments by freezing live LiDAR streams into traceable snapshots for interactive inspection and on-demand mesh generation within an immersive VR setting.
The paper proposes SMPD, a novel self-supervised meta-learning framework that leverages multi-modal temporal data and contrastive pre-training to predict IoT device compromises 24–48 hours in advance with high accuracy and few-shot adaptability, significantly outperforming existing intrusion detection baselines in resource-constrained environments.
This paper introduces GRADV, a reproducible signal-analysis workflow that discovers and evaluates target-class audio patterns in keyword-spotting models by optimizing waveforms and performing segment-level analysis, achieving high success rates without proposing new speech-recognition architectures.
This paper proposes a Pydantic schema and a few-shot small language model baseline to extract engineering change traceability from fragmented, bilingual, and truncated legacy CNC work orders, demonstrating high schema validity while revealing that contextually grounded pre-prompts degrade performance and deterministic post-filters improve precision without affecting recall.
OptiVerse AI is an explainable deep learning system that utilizes EfficientNetB7 and Grad-CAM to classify diabetic retinopathy and reconstruct 3D retinal structures from standard fundus images, achieving over 95% accuracy to enhance early disease detection and clinical decision-making.
This paper introduces RECON, a neuro-symbolic multi-agent architecture that combines deterministic rules, fuzzy detection, and LLM reasoning with a YAML-driven recipe system and tamper-evident auditing to deliver reproducible, defensible, and evidence-fused document compliance reviews across heterogeneous enterprise repositories.
This paper demonstrates that for correlative microscopy registration, cross-modal appearance similarity is a more critical determinant of success than field-of-view scale handling, as evidenced by the superior performance of cross-modal-trained backbones and the failure of naive scale-aware wrappers to overcome appearance mismatches.