arXiv subjects

Xiang Zhai

Publications and source records attributed to Xiang Zhai.

7 recordsLinked to original sources

arXiv2026

Bridging Textual Profiles and Latent User Embeddings for Personalization

Personalized systems rely on user representations to connect behavioral history with downstream recommendation applications. Existing methods typically employ either supervised latent user embeddings, which are effective for retrieval but difficult to interpret, or textual user profiles, which are interpretable but challenging to optimize for downstream utility due to lack of direct supervision. To bridge this gap, we present BLUE, a reinforcement learning framework that unifies these two forms of user representation by aligning language-based user profiles with embedding-based recommendation objectives. Given a user interaction history, BLUE leverages a profiler Large Language Model (LLM) to generate textual profiles, while an embedding model provides reward signals. This encourages the resulting textual representations to move closer to positive items and farther from negative ones in the embedding space. We further introduce a text-space supervision signal based on next-item prediction, ensuring the learned profiles remain both semantically meaningful and highly effective for downstream retrieval. Experiments on Amazon Reviews 2023 and Google Local Reviews in zero-shot sequential recommendation settings demonstrate that BLUE consistently outperforms strong baselines under both frozen and trainable embedding conditions. Notably, BLUE achieves clear gains in cross-domain transfer, highlighting the strong generalization ability of the learned user profiles. Furthermore, these generated profiles provide superior personalized context for question answering compared to raw user histories or alternative profile optimization methods. Overall, these results show that BLUE provides an effective way to unify interpretable textual profiling with discriminative latent embeddings for personalization.

cs.IR↗

arXiv2025

Efficient Third-Harmonic Generation via Strong Coupling of Quasi-Guided Modes

Improving the conversion efficiency is critical for advancing nonlinear photonic devices, enabling applications in ultrafast optics, quantum light sources, and high-resolution imaging. Herein, we present a bilayer waveguide integrated with a periodic square nanocube array to enhance third-harmonic generation (THG) efficiency. This structure achieves strong coupling between TE- and TM-polarized quasi-guided mode (QGM) resonances, thereby enabling simultaneous dynamic control of both resonant Q-factor and coupling strength through incident polarization angle. The distinct avoided crossing observed in both reflection spectra and eigenfrequency diagrams indicates the strong coupling regime. This mechanism, facilitated by the inter-mode energy exchange within the hybridized system, results in an enhancement of Q-factor by two orders of magnitude when compared to conventional guided-mode resonances. Simulation results verify an unprecedented maximum Q-factor of 10^12 for the upper-branch mode. Leveraging strong-coupling-induced field enhancement, the system attains THG conversion efficiency of order 10^-2, demonstrating unprecedented nonlinear performance.

physics.optics↗

arXiv2024

CALRec: Contrastive Alignment of Generative LLMs for Sequential Recommendation

Traditional recommender systems such as matrix factorization methods have primarily focused on learning a shared dense embedding space to represent both items and user preferences. Subsequently, sequence models such as RNN, GRUs, and, recently, Transformers have emerged and excelled in the task of sequential recommendation. This task requires understanding the sequential structure present in users' historical interactions to predict the next item they may like. Building upon the success of Large Language Models (LLMs) in a variety of tasks, researchers have recently explored using LLMs that are pretrained on vast corpora of text for sequential recommendation. To use LLMs for sequential recommendation, both the history of user interactions and the model's prediction of the next item are expressed in text form. We propose CALRec, a two-stage LLM finetuning framework that finetunes a pretrained LLM in a two-tower fashion using a mixture of two contrastive losses and a language modeling loss: the LLM is first finetuned on a data mixture from multiple domains followed by another round of target domain finetuning. Our model significantly outperforms many state-of-the-art baselines (+37% in Recall@1 and +24% in NDCG@10) and our systematic ablation studies reveal that (i) both stages of finetuning are crucial, and, when combined, we achieve improved performance, and (ii) contrastive alignment is effective among the target domains explored in our experiments.

cs.IR↗

arXiv2020

Fast Modeling and Understanding Fluid Dynamics Systems with Encoder-Decoder Networks

Is a deep learning model capable of understanding systems governed by certain first principle laws by only observing the system's output? Can deep learning learn the underlying physics and honor the physics when making predictions? The answers are both positive. In an effort to simulate two-dimensional subsurface fluid dynamics in porous media, we found that an accurate deep-learning-based proxy model can be taught efficiently by a computationally expensive finite-volume-based simulator. We pose the problem as an image-to-image regression, running the simulator with different input parameters to furnish a synthetic training dataset upon which we fit the deep learning models. Since the data is spatiotemporal, we compare the performance of two alternative treatments of time; a convolutional LSTM versus an autoencoder network that treats time as a direct input. Adversarial methods are adopted to address the sharp spatial gradient in the fluid dynamic problems. Compared to traditional simulation, the proposed deep learning approach enables much faster forward computation, which allows us to explore more scenarios with a much larger parameter space given the same time. It is shown that the improved forward computation efficiency is particularly valuable in solving inversion problems, where the physics model has unknown parameters to be determined by history matching. By computing the pixel-level attention of the trained model, we quantify the sensitivity of the deep learning model to key physical parameters and hence demonstrate that the inversion problems can be solved with great acceleration. We assess the efficacy of the machine learning surrogate in terms of its training speed and accuracy. The network can be trained within minutes using limited training data and achieve accuracy that scales desirably with the amount of training data supplied.

physics.comp-ph↗

arXiv2014

Three-Dimensional MHD Simulation of Caltech Plasma Jet Experiment: First Results

Magnetic fields are believed to play an essential role in astrophysical jets with observations suggesting the presence of helical magnetic fields. Here, we present three-dimensional (3D) ideal MHD simulationsof the Caltech plasma jet experiment using a magnetic tower scenario as the baseline model. Magnetic fields consist of an initially localized dipole-like poloidal component and a toroidal component that is continuously being injected into the domain. This flux injection mimics the poloidal currents driven by the anode-cathode voltage drop in the experiment. The injected toroidal field stretches the poloidal fields to large distances, while forming a collimated jet along with several other key features. Detailed comparisons between 3D MHD simulations and experimental measurements provide a comprehensive description of the interplay among magnetic force, pressure and flow effects. In particular, we delineate both the jet structure and the transition process that converts the injected magnetic energy to other forms. With suitably chosen parameters that are derived from experiments, the jet in the simulation agrees quantitatively with the experimental jet in terms of magnetic/kinetic/inertial energy, total poloidal current, voltage, jet radius, and jet propagation velocity. Specifically, the jet velocity in the simulation is proportional to the poloidal current divided by the square root of the jet density, in agreement with both the experiment and analytical theory. This work provides a new and quantitative method for relating experiments, numerical simulations and astrophysical observation, and demonstrates the possibility of using terrestrial laboratory experiments to study astrophysical jets.

astro-ph.HE↗

arXiv2010

Diffuse X-ray Emissions from Dynamic Planetary Nebulae

We present theoretical results of a piecewise isothermal shock wind model devised for predicting the luminosity and surface brightness profile of diffuse X-ray emissions primarily from the inner shocked downstream wind zone of a planetary nebula (PN) surrounded by self-similar shocked dense shell and outer slow AGB wind envelope involving self-gravity and compare/fit our computational model results with available observations of a few grossly spherical X-ray emitting PNe. Matching shocked piecewise isothermal self-similar void (ISSV) solutions with self-gravity of Lou & Zhai (LZ) for the outer zone and a stationary isothermal fast tenuous wind with a reverse shock for the inner zone across an expanding contact discontinuity, we can consistently construct dynamic evolution models of PNe with diffuse X-ray emissions. On the basis of such a chosen dynamic wind interaction model, both X-ray luminosity and radial X-ray brightness profile are determined by three key parameters, namely the so-called X-ray parameter $X$, two radii $R_{rs}$ and $R_c$ of the reverse shock and the contact discontinuity. We find that morphologies of X-ray emissions would appear in the forms of either a central luminous sphere or a bright ring embedded within optically bright shells. In contrast to previous adiabatic models, the X-ray brightness peaks around the reverse shock, instead of the contact discontinuity surface just inside the outer shocked dense shell. Diffuse X-ray emissions of a few observed PNe appear to support this wind-wind dynamic interaction scenario.

astro-ph.SR↗

arXiv2009

Dynamic Evolution Model of Isothermal Voids and Shocks

We explore self-similar hydrodynamic evolution of central voids embedded in an isothermal gas of spherical symmetry under the self-gravity. More specifically, we study voids expanding at constant radial speeds in an isothermal gas and construct all types of possible void solutions without or with shocks in surrounding envelopes. We examine properties of void boundaries and outer envelopes. Voids without shocks are all bounded by overdense shells and either inflows or outflows in the outer envelope may occur. These solutions, referred to as type $\mathcal{X}$ void solutions, are further divided into subtypes $\mathcal{X}_{\rm I}$ and $\mathcal{X}_{\rm II}$ according to their characteristic behaviours across the sonic critical line (SCL). Void solutions with shocks in envelopes are referred to as type $\mathcal{Z}$ voids and can have both dense and quasi-smooth edges. Asymptotically, outflows, breezes, inflows, accretions and static outer envelopes may all surround such type $\mathcal{Z}$ voids. Both cases of constant and varying temperatures across isothermal shock fronts are analyzed; they are referred to as types $\mathcal{Z}_{\rm I}$ and $\mathcal{Z}_{\rm II}$ void shock solutions. We apply the `phase net matching procedure' to construct various self-similar void solutions. We also present analysis on void generation mechanisms and describe several astrophysical applications. By including self-gravity, gas pressure and shocks, our isothermal self-similar void (ISSV) model is adaptable to various astrophysical systems such as planetary nebulae, hot bubbles and superbubbles in the interstellar medium as well as supernova remnants.

astro-ph.SR↗