arXiv subjects

Jingyi Shen

Publications and source records attributed to Jingyi Shen.

8 recordsLinked to original sources

arXiv2026

HydraHead: From Head-Level Functional Heterogeneity to Specialized Attention Hybridization

The quadratic complexity of attention poses a critical bottleneck for long-context processing, spurring interest in hybrid attention designs. Most open-source hybrid models adopt a layer-wise strategy. Yet, prior work has noted the inherent difficulty of integrating Linear Attention (LA) with Full Attention (FA), suggesting that the design space of attention hybridization remains underexplored. To probe this space, we conduct interpretability analysis and observe that layers exhibit block-wise functional similarity, while individual heads within the same layer display distinct functional specialization despite sharing input features. This head-level heterogeneity suggests that the head dimension provides a natural and principled granularity for fusing heterogeneous attention signals. Building on this insight, we introduce HydraHead, a novel architecture that hybridizes FA and LA along the head axis. HydraHead features two key innovations: (1) an interpretability-driven selection strategy that identifies retrieval-critical heads and preserves FA only for them, and (2) a scale-normalized fusion module that reconciles the distributional gap between FA and LA head outputs. By leveraging a three-stage transfer pipeline with parameter reuse and distillation, we achieve high-performance hybrid models with minimal training overhead. Under a unified training setup, HydraHead outperforms other hybrid designs in long-context tasks while maintaining strong general reasoning. With interpretability-driven head selection, it matches a 3:1 layer-wise hybrid's long-context performance at a 7:1 LA-to-FA ratio. Crucially, trained on only 15B tokens, HydraHead achieves over 69% improvement over the baseline at 512K context length, approaching Qwen3.5, a leading model of comparable size with a native context length of 256K. This highlights the significant scaling potential of head-level hybridization.

cs.CL↗

arXiv2026

SOLARIS: Speculative Offloading of Latent-bAsed Representation for Inference Scaling

Recent advances in recommendation scaling laws have led to foundation models of unprecedented complexity. While these models offer superior performance, their computational demands make real-time serving impractical, often forcing practitioners to rely on knowledge distillation-compromising serving quality for efficiency. To address this challenge, we present SOLARIS (Speculative Offloading of Latent-bAsed Representation for Inference Scaling), a novel framework inspired by speculative decoding. SOLARIS proactively precomputes user-item interaction embeddings by predicting which user-item pairs are likely to appear in future requests, and asynchronously generating their foundation model representations ahead of time. This approach decouples the costly foundation model inference from the latency-critical serving path, enabling real-time knowledge transfer from models previously considered too expensive for online use. Deployed across Meta's advertising system serving billions of daily requests, SOLARIS achieves 0.67% revenue-driving top-line metrics gain, demonstrating its effectiveness at scale.

cs.LG↗

arXiv2025

Intertwined nematic and d-wave superconductive orders in optimally-doped La1.84Sr0.16CuO4

The anisotropy of the superconducting state and superconducting fluctuations in the CuO2 plane is directly related to the superconducting mechanism of copper oxide superconductors and is therefore pivotal for understanding high-temperature superconductivity. Here, we integrated the high-precision angle-resolved resistivity (ARR) measurement with a rotatable in-plane magnetic field to systematically study the angular dependence of superconducting fluctuations in optimally doped La1.84Sr0.16CuO4 (LSCO). By independently controlling the directions of the current and the magnetic field, we are able to isolate the magneto-resistivity contributed by the superconducting vortex motion and distinguish excitations from nematic superconductivity and d-wave superconductive order based on their respective C2 and C4 symmetries. Signatures of two intertwined superconductive orders are also evident in the measured angular dependence of the critical current. A T-B phase diagram of different types of superconducting fluctuations is determined. These findings are closely related to other intriguing phenomena, such as pair density wave and charge density wave.

cond-mat.supr-con↗

arXiv2025

Air-stable lithiation engineering of $\mathrm{MoS}_{2}$ for direct-bandgap multilayers

Due to its sizable direct bandgap and strong light-matter interactions, the preparation of monolayer $\mathrm{MoS}_{2}$ has attracted significant attention and intensive research efforts. However, multilayer $\mathrm{MoS}_{2}$ is largely overlooked because of its optically inactive indirect bandgap caused by interlayer coupling. It is highly desirable to modulate and decrease the interlayer coupling so that each layer in multilayer $\mathrm{MoS}_{2}$ can exhibit a monolayer-like direct-gap behavior. Here, we demonstrate the nanoprobe fabrication of $\mathrm{Li}_{x}\mathrm{MoS}_{2}$-based multilayers exhibiting a direct bandgap and strong photoluminescence emission from tightly bound excitons and trions. The fabrication is facilitated by our newly developed Li-ion platform, featuring tip-induced Li intercalation, air stability and rewritability. Raman characterizations reveal that controlled Li intercalation effectively transforms multilayer $\mathrm{MoS}_{2}$ into the stack of multiple monolayers, leading to a 26-fold enhancement of photoluminescence, compared to a monolayer. This intercalation result is different from existing observations of transforming $\mathrm{MoS}_{2}$ multilayers into metallic phases.

cond-mat.mes-hall↗

arXiv2024

SurroFlow: A Flow-Based Surrogate Model for Parameter Space Exploration and Uncertainty Quantification

Existing deep learning-based surrogate models facilitate efficient data generation, but fall short in uncertainty quantification, efficient parameter space exploration, and reverse prediction. In our work, we introduce SurroFlow, a novel normalizing flow-based surrogate model, to learn the invertible transformation between simulation parameters and simulation outputs. The model not only allows accurate predictions of simulation outcomes for a given simulation parameter but also supports uncertainty quantification in the data generation process. Additionally, it enables efficient simulation parameter recommendation and exploration. We integrate SurroFlow and a genetic algorithm as the backend of a visual interface to support effective user-guided ensemble simulation exploration and visualization. Our framework significantly reduces the computational costs while enhancing the reliability and exploration capabilities of scientific surrogate models.

cs.LG↗

arXiv2023

PSRFlow: Probabilistic Super Resolution with Flow-Based Models for Scientific Data

Although many deep-learning-based super-resolution approaches have been proposed in recent years, because no ground truth is available in the inference stage, few can quantify the errors and uncertainties of the super-resolved results. For scientific visualization applications, however, conveying uncertainties of the results to scientists is crucial to avoid generating misleading or incorrect information. In this paper, we propose PSRFlow, a novel normalizing flow-based generative model for scientific data super-resolution that incorporates uncertainty quantification into the super-resolution process. PSRFlow learns the conditional distribution of the high-resolution data based on the low-resolution counterpart. By sampling from a Gaussian latent space that captures the missing information in the high-resolution data, one can generate different plausible super-resolution outputs. The efficient sampling in the Gaussian latent space allows our model to perform uncertainty quantification for the super-resolved results. During model training, we augment the training data with samples across various scales to make the model adaptable to data of different scales, achieving flexible super-resolution for a given input. Our results demonstrate superior performance and robust uncertainty quantification compared with existing methods such as interpolation and GAN-based super-resolution networks.

eess.IV↗

arXiv2022

IDLat: An Importance-Driven Latent Generation Method for Scientific Data

Deep learning based latent representations have been widely used for numerous scientific visualization applications such as isosurface similarity analysis, volume rendering, flow field synthesis, and data reduction, just to name a few. However, existing latent representations are mostly generated from raw data in an unsupervised manner, which makes it difficult to incorporate domain interest to control the size of the latent representations and the quality of the reconstructed data. In this paper, we present a novel importance-driven latent representation to facilitate domain-interest-guided scientific data visualization and analysis. We utilize spatial importance maps to represent various scientific interests and take them as the input to a feature transformation network to guide latent generation. We further reduced the latent size by a lossless entropy encoding algorithm trained together with the autoencoder, improving the storage and memory efficiency. We qualitatively and quantitatively evaluate the effectiveness and efficiency of latent representations generated by our method with data from multiple scientific visualization applications.

cs.LG↗

arXiv2020

An Information-theoretic Visual Analysis Framework for Convolutional Neural Networks

Despite the great success of Convolutional Neural Networks (CNNs) in Computer Vision and Natural Language Processing, the working mechanism behind CNNs is still under extensive discussions and research. Driven by a strong demand for the theoretical explanation of neural networks, some researchers utilize information theory to provide insight into the black box model. However, to the best of our knowledge, employing information theory to quantitatively analyze and qualitatively visualize neural networks has not been extensively studied in the visualization community. In this paper, we combine information entropies and visualization techniques to shed light on how CNN works. Specifically, we first introduce a data model to organize the data that can be extracted from CNN models. Then we propose two ways to calculate entropy under different circumstances. To provide a fundamental understanding of the basic building blocks of CNNs (e.g., convolutional layers, pooling layers, normalization layers) from an information-theoretic perspective, we develop a visual analysis system, CNNSlicer. CNNSlicer allows users to interactively explore the amount of information changes inside the model. With case studies on the widely used benchmark datasets (MNIST and CIFAR-10), we demonstrate the effectiveness of our system in opening the blackbox of CNNs.

cs.CV↗