Rethinking Small VLM Quantization: From Component-Wise Analysis to Hardware-Aware Edge Deployment

The emergence of vision language models with fewer than 3 billion parameters has accelerated the implementation of on-device multimodal intelligence. However, a detailed understanding of component-wise quantization remains a bottleneck for optimal deployment. This paper presents a systematic evaluation framework for empirically validating five hypotheses across six quantization configurations on the Jetson Orin NX and AGX. By separating the vision encoder, projector, and large language model backbone yields the following results: (1) Quantization sensitivity is governed by the structural paradigm (MoE vs. dense) rather than scale alone, with MoE backbones mitigating INT4 noise where dense backbones degrade; (2) SigLIP encoders incur disproportionate INT8 latency on Jetson Ampere--a deployment-specific encoder-kernel-hardware interaction, not a SigLIP flaw; (3) Although INT4 quantization of LLMs greatly reduces VRAM consumption, it also causes slower token generation due to dequantization overhead; (4) Composite quantization errors are largely additive, except along the modality-alignment path, which is architecture-dependent; (5) The intelligence-per-joule profile varies significantly across platforms owing to memory bandwidth constraints.

Spectral and Temporal Denoising for Differentially Private Optimization

This paper introduces the FFT-Enhanced Kalman Filter (FFTKF), a differentially private optimization method that addresses the challenge of preserving performance in DP-SGD, where added noise typically degrades model utility. FFTKF integrates frequency-domain noise shaping with Kalman filtering to enhance gradient quality while preserving $(\varepsilon, \delta)$-DP guarantees. It employs a high-frequency shaping mask in the Fourier domain to concentrate differential privacy noise in less informative spectral components, preserving low-frequency gradient signals. A scalar-gain Kalman filter with finite-difference Hessian approximation further refines the denoised gradients. With a per-iteration complexity of $\mathcal{O}(d \log d)$, FFTKF demonstrates improved test accuracy over DP-SGD and DiSK across MNIST, CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets using CNNs, Wide ResNets, and Vision Transformers. Theoretical analysis confirms that FFTKF maintains equivalent privacy guarantees while achieving a tighter privacy-utility trade-off through reduced noise and controlled bias.

Improving the Real-Data Driven Network Evaluation Model for Digital Twin Networks

With the emergence and proliferation of new forms of large-scale services such as smart homes, virtual reality/augmented reality, the increasingly complex networks are raising concerns about significant operational costs. As a result, the need for network management automation is emphasized, and Digital Twin Networks (DTN) technology is expected to become the foundation technology for autonomous networks. DTN has the advantage of being able to operate and system networks based on real-time collected data in a closed-loop system, and currently it is mainly designed for optimization scenarios. To improve network performance in optimization scenarios, it is necessary to select appropriate configurations and perform accurate performance evaluation based on real data. However, most network evaluation models currently use simulation data. Meanwhile, according to DTN standards documents, artificial intelligence (AI) models can ensure scalability, real-time performance, and accuracy in large-scale networks. Various AI research and standardization work is ongoing to optimize the use of DTN. When designing AI models, it is crucial to consider the characteristics of the data. This paper presents an autoencoder-based skip connected message passing neural network (AE-SMPN) as a network evaluation model using real network data. The model is created by utilizing graph neural network (GNN) with recurrent neural network (RNN) models to capture the spatiotemporal features of network data. Additionally, an AutoEncoder (AE) is employed to extract initial features. The neural network was trained using the real DTN dataset provided by the Barcelona Neural Networking Center (BNN-UPC), and the paper presents the analysis of the model structure along with experimental results.