arXiv subjects

Yiying Dong

Publications and source records attributed to Yiying Dong.

5 recordsLinked to original sources

arXiv2026

GPrune-LLM: Generalization-Aware Structured Pruning for Large Language Models

Structured pruning is widely applied to compress large language models (LLMs), but its performance depends heavily on how neuron importance is estimated. Most existing methods rely on activation statistics from a single calibration set, which introduces calibration bias and degrades downstream cross-task generalization. We observe that neurons exhibit heterogeneous distribution sensitivity, ranging from maintaining relatively stable rankings across calibration datasets to showing substantially larger cross-dataset variation. Ignoring this heterogeneity, existing methods rank all neurons in shared spaces with a uniform scoring source, so calibration-specific neurons dominate the ranking and weakly-activated neurons are scored unreliably. To address this, we propose GPrune-LLM, a structured pruning framework that reduces calibration bias by measuring and exploiting the cross-distribution behavior of neurons for fair comparison. Specifically, we restructure the neuron ranking space into behavior-consistent local spaces, adapt the scoring source where the calibration signal is unreliable, and learn per-module sparsity allocation under a global budget. Experiments across multiple models and downstream tasks show that GPrune-LLM improves the generalization of its base pruning metrics, with gains most pronounced at high sparsity, and reduces dependence on the choice of importance metric.

cs.LG↗

arXiv2026

Programmable electro-optic frequency comb empowers integrated parallel convolution processing

Integrated photonic convolution processors make optical neural networks (ONNs) a transformative solution for artificial intelligence applications such as machine vision. To enhance the parallelism, throughput, and energy efficiency of ONNs, wavelength multiplexing is widely applied. However, it often encounters the challenges of low compactness, limited scalability, and high weight reconstruction latency. Here, we proposed and demonstrated an integrated photonic processing unit with a parallel convolution computing speed of 1.62 trillion operations per second (TOPS) and a weight reconstruction speed exceeding 38 GHz. This processing unit simultaneously achieves, for the first time, multi-wavelength generation and weight mapping via a single programmable electro-optic (EO) frequency comb, featuring unprecedented compactness, device-footprint independent scalability, and near-unity optical power conversion efficiency (conversion efficiency from input optical power to output weighted comb lines). To demonstrate the reconfigurability and functionality of this processing unit, we implemented image edge detection and object classification based on EO combs obtained using the particle swarm algorithm and an EO comb neural network training framework, respectively. Our programmable EO comb-based processing framework establishes a new paradigm towards the development of low-latency monolithic photonic processors, promising real-time in-sensor learning for autonomous vehicles, intelligent robotics, and drones.

physics.optics↗

arXiv2026

REF-VLM: Triplet-Based Referring Paradigm for Unified Visual Decoding

Multimodal Large Language Models (MLLMs) demonstrate robust zero-shot capabilities across diverse vision-language tasks after training on mega-scale datasets. However, dense prediction tasks, such as semantic segmentation and keypoint detection, pose significant challenges for MLLMs when represented solely as text outputs. Simultaneously, current MLLMs utilizing latent embeddings for visual task decoding generally demonstrate limited adaptability to both multi-task learning and multi-granularity scenarios. In this work, we present \textbf{REF-VLM}, an end-to-end framework for unified training of various visual decoding tasks. To address complex visual decoding scenarios, we introduce the \textbf{Triplet-Based Referring Paradigm (TRP)}, which explicitly decouples three critical dimensions in visual decoding tasks through a triplet structure: concepts, decoding types, and targets. TRP employs symbolic delimiters to enforce structured representation learning, enhancing the parsability and interpretability of model outputs. Additionally, we construct \textbf{Visual-Task Instruction Following Dataset (VT-Instruct)}, a large-scale multi-task dataset containing over 100 million multimodal dialogue samples across 25 task types. Beyond text inputs and outputs, VT-Instruct incorporates various visual prompts such as point, box, scribble, and mask, and generates outputs composed of text and visual units like box, keypoint, depth and mask. The combination of different visual prompts and visual units generates a wide variety of task types, expanding the applicability of REF-VLM significantly. Both qualitative and quantitative experiments demonstrate that our REF-VLM outperforms other MLLMs across a variety of standard benchmarks. The code, dataset, and demo will be publicly available.

cs.CV↗

arXiv2025

Scalable multilayer diffractive neural network with all-optical nonlinear activation

All-optical diffractive neural networks (DNNs) offer a promising alternative to electronics-based neural network processing due to their low latency, high throughput, and inherent spatial parallelism. However, the lack of reconfigurability and nonlinearity limits existing all-optical DNNs to handling only simple tasks. In this study, we present a folded optical system that enables a multilayer reconfigurable DNN using a single spatial light modulator. This platform not only enables dynamic weight reconfiguration for diverse classification challenges but crucially integrates a mirror-coated silicon substrate exhibiting instantaneous \c{hi}(3) nonlinearity. The incorporation of all-optical nonlinear activation yields substantial accuracy improvements across benchmark tasks, with performance gains becoming increasingly significant as both network depth and task complexity escalate. Our system represents a critical advancement toward realizing scalable all-optical neural networks with complex architectures, potentially achieving computational capabilities that rival their electronic counterparts while maintaining photonic advantages.

physics.optics↗

arXiv2024

Metasurface-generated large and arbitrary analog convolution kernels for accelerated machine vision

In the rapidly evolving field of artificial intelligence, convolutional neural networks are essential for tackling complex challenges such as machine vision and medical diagnosis. Recently, to address the challenges in processing speed and power consumption of conventional digital convolution operations, many optical components have been suggested to replace the digital convolution layer in the neural network, accelerating various machine vision tasks. Nonetheless, the analog nature of the optical convolution kernel has not been fully explored. Here, we develop a spatial frequency domain training method to create arbitrarily shaped analog convolution kernels using an optical metasurface as the convolution layer, with its receptive field largely surpassing digital convolution kernels. By employing spatial multiplexing, the multiple parallel convolution kernels with both positive and negative weights are generated under the incoherent illumination condition. We experimentally demonstrate a 98.59% classification accuracy on the MNIST dataset, with simulations showing 92.63% and 68.67% accuracy on the Fashion-MNIST and CIFAR-10 datasets with additional digital layers. This work underscores the unique advantage of analog optical convolution, offering a promising avenue to accelerate machine vision tasks, especially in edge devices.

physics.optics↗