arXiv subjects

Weijie Huang

Publications and source records attributed to Weijie Huang.

18 recordsLinked to original sources

arXiv2026

Hoss: Fast Oblivious Semantic Search with Heterogeneous GPU-CPU-TEE Architecture

Semantic search is widely deployed in modern AI systems, but protecting both data contents and access patterns remains challenging. The current state-of-the-art system, Compass, achieves oblivious semantic search by building an optimized ORAM over HNSW graphs. However, even with aggressive optimizations, it still incurs large overheads. Closing this performance gap is fundamentally difficult: Compass has already removed most cryptographic overheads, leaving ORAM accesses as the dominant cost, which are constrained by well-known Omega(log N) bandwidth lower bounds. Our key insight is that traditional ORAM overhead stems from the assumption of limited private memory, whereas modern GPU TEEs provide large private memory (Pmem) that blinds internal access patterns (Hunt et al., NSDI '23). This shift opens a new design space. We therefore propose Hoss, a first-of-its-kind oblivious semantic search system with a heterogeneous CPU-GPU TEE architecture that supports fast, scalable search with low cost of ownership. In Hoss, the GPU TEE's large Pmem hosts the hot-path HNSW traversal, while the lower layers of the graph, if they exceed GPU capacity, are offloaded to CPU TEEs. The system invokes oblivious primitives only when accessing these lower layers. The availability of large Pmem also enables new optimization opportunities. For example, Hoss features a host-access ORAM mechanism that goes beyond traditional performance constraints and incorporates several data-dependent optimizations that are not possible in prior designs. We implement a prototype of Hoss and benchmark it against Compass. Our results show that Hoss achieves up to 67x speedup while maintaining high recall, with larger gains at scale.

cs.CR↗

arXiv2026

An Onsager Variational Scheme for Pressure-Driven Tumor Growth and Hele-Shaw Limits

Pressure-driven tumor growth models describe the coupling between cell proliferation and mechanical pressure and naturally lead to moving free boundary problems. Their numerical approximation is challenging due to degenerate diffusion, pressure-dependent proliferation, and the stiffness of the pressure law $p=n^γ$ for large $γ$. In this paper, we propose a structure-preserving finite difference method for this class of pressure-driven tumor growth models with pressure-dependent proliferation. The method is derived from the Onsager variational principle. The key idea is to introduce a modified energy shifted by the homeostatic pressure, so that the growth term can be written in a dissipative form and incorporated together with the transport part into a unified Rayleighian formulation. This formulation leads to a time-discrete constrained minimization problem and a fully discrete scheme with explicit mobilities and an implicit pressure update. We prove that the scheme preserves nonnegativity and the homeostatic upper bound, satisfies a discrete modified energy dissipation law, and admits a fixed-grid stiff-pressure limiting structure. Numerical experiments in one and two spatial dimensions demonstrate the accuracy of the method, its convergence toward the Hele--Shaw limit for large $γ$, and its ability to capture free boundary evolutions with topology changes.

math.NA↗

arXiv2026

A WKB-based fixed-grid method for capturing trait concentration in a dispersal evolution model

The evolution of dispersal traits is a fundamental topic in evolutionary ecology, where natural selection may drive the trait distribution toward concentration in the rare-mutation regime. This singular behavior poses a serious numerical difficulty, since direct discretizations of the population density require very fine trait grids to identify the fittest trait and to resolve the concentrated profile accurately. In this paper, we develop a WKB-based numerical framework for a dispersal evolution model. By separating the exponentially concentrated trait dependence from a smoother amplitude variable and combining this WKB representation with dual trait-grid implementation and other specially designed techniques, the method recovers the selected trait and the associated concentration structure accurately and efficiently on fixed trait grids. We establish a semi-discrete stationary fixed-grid asymptotic-preserving structure for the rare-mutation limit of the steady-state problem. Numerical experiments compare the proposed method with direct density discretizations and confirm its advantage in the small-mutation regime.

math.NA↗

arXiv2026

Ground states and droplet regimes of the extended Gross-Pitaevskii equation with Lee-Huang-Yang correction

We study the ground states of the extended Gross--Pitaevskii equation with the Lee--Huang--Yang correction from both theoretical and numerical perspectives. Starting from the three-dimensional model, we derive reduced one- and two-dimensional equations through nondimensionalization and dimensional reduction. We establish existence and nonexistence results for ground states in different spatial dimensions, both in free space and under confining external potentials. For the numerical computation of ground states, we propose a normalized gradient flow method with a Lagrange multiplier. The numerical results show how the model parameters affect the ground-state profiles, and reveal different regimes in the free-space parameter plane, including no-ground-state, soliton-like, and droplet-like regions. We also introduce a simple flat-top approximation for the droplet regime and present two- and three-dimensional computations to illustrate more general localized structures.

math-ph↗

arXiv2026

A sharp-interface model for solid-state dewetting with wetting potential

We propose a sharp-interface model for solid-state dewetting of thin films with wetting potential, where the wetting effect is incorporated through a thickness-dependent surface energy. The model is governed by surface diffusion together with natural boundary conditions, and describes the morphological evolution of the film-vapor interface. For its numerical approximation, we develop an efficient semi-implicit finite element method based on a Taylor expansion of the wetting-potential term. Numerical simulations in two dimensions show that the proposed model and method can capture various dewetting phenomena. They also indicate that, as the range of the wetting potential tends to zero, the proposed model approaches the sharp-interface model with thickness-independent surface energy proposed in [1]. The model and numerical method are further extended to three dimensions, where the computations capture complex morphological evolution in solid-state dewetting.

math.NA↗

arXiv2026

Large Language Model-Powered Evolutionary Code Optimization on a Phylogenetic Tree

Optimizing scientific computing algorithms for modern GPUs is a labor-intensive and iterative process involving repeated code modification, benchmarking, and tuning across complex hardware and software stacks. Recent work has explored large language model (LLM)-assisted evolutionary methods for automated code optimization, but these approaches primarily rely on outcome-based selection and random mutation, underutilizing the rich trajectory information generated during iterative optimization. We propose PhyloEvolve, an LLM-agent system that reframes GPU-oriented algorithm optimization as an In-Context Reinforcement Learning (ICRL) problem. This formulation enables trajectory-conditioned reuse of optimization experience without model retraining. PhyloEvolve integrates Algorithm Distillation and prompt-based Decision Transformers into an iterative workflow, treating sequences of algorithm modifications and performance feedback as first-class learning signals. To organize optimization history, we introduce a phylogenetic tree representation that captures inheritance, divergence, and recombination among algorithm variants, enabling backtracking, cross-lineage transfer, and reproducibility. The system combines elite trajectory pooling, multi-island parallel exploration, and containerized execution to balance exploration and exploitation across heterogeneous hardware. We evaluate PhyloEvolve on scientific computing workloads including PDE solvers, manifold learning, and spectral graph algorithms, demonstrating consistent improvements in runtime, memory efficiency, and correctness over baseline and evolutionary methods. Code is published at: https://github.com/annihi1ation/phylo_evolve

cs.AI↗

arXiv2025

Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking

Large Language Models (LLMs) have demonstrated notable capabilities across financial tasks, including financial report summarization, earnings call transcript analysis, and asset classification. However, their real-world effectiveness in managing complex fund investment remains inadequately assessed. A fundamental limitation of existing benchmarks for evaluating LLM-driven trading strategies is their reliance on historical back-testing, inadvertently enabling LLMs to "time travel"-leveraging future information embedded in their training corpora, thus resulting in possible information leakage and overly optimistic performance estimates. To address this issue, we introduce DeepFund, a live fund benchmark tool designed to rigorously evaluate LLM in real-time market conditions. Utilizing a multi-agent architecture, DeepFund connects directly with real-time stock market data-specifically data published after each model pretraining cutoff-to ensure fair and leakage-free evaluations. Empirical tests on nine flagship LLMs from leading global institutions across multiple investment dimensions-including ticker-level analysis, investment decision-making, portfolio management, and risk control-reveal significant practical challenges. Notably, even cutting-edge models such as DeepSeek-V3 and Claude-3.7-Sonnet incur net trading losses within DeepFund real-time evaluation environment, underscoring the present limitations of LLMs for active fund management. Our code is available at https://github.com/HKUSTDial/DeepFund.

cs.CE↗

arXiv2024

AdaCo: Overcoming Visual Foundation Model Noise in 3D Semantic Segmentation via Adaptive Label Correction

Recently, Visual Foundation Models (VFMs) have shown a remarkable generalization performance in 3D perception tasks. However, their effectiveness in large-scale outdoor datasets remains constrained by the scarcity of accurate supervision signals, the extensive noise caused by variable outdoor conditions, and the abundance of unknown objects. In this work, we propose a novel label-free learning method, Adaptive Label Correction (AdaCo), for 3D semantic segmentation. AdaCo first introduces the Cross-modal Label Generation Module (CLGM), providing cross-modal supervision with the formidable interpretive capabilities of the VFMs. Subsequently, AdaCo incorporates the Adaptive Noise Corrector (ANC), updating and adjusting the noisy samples within this supervision iteratively during training. Moreover, we develop an Adaptive Robust Loss (ARL) function to modulate each sample's sensitivity to noisy supervision, preventing potential underfitting issues associated with robust loss. Our proposed AdaCo can effectively mitigate the performance limitations of label-free learning networks in 3D semantic segmentation tasks. Extensive experiments on two outdoor benchmark datasets highlight the superior performance of our method.

cs.CV↗

arXiv2024

A parameter-free statistical model for two-dimensional carbon nanostructures

Energy degeneracy in physical systems may be induced by symmetries of the Hamiltonian, and the resonance of degeneracy states in carbon nanostructures can effectively enhance the stability of the system. Combining the octet rule, we introduce a parameter-free statistical model to determine the physical properties by lifting the energy degeneracy in carbon nanostructures. This model offers a direct path to accurately ascertain electron density distributions in quantum systems, akin to how charge density is used in density functional theory to deduce system properties. Our methodology diverges from traditional quantum mechanics, focusing instead on this unique statistical model by minimizing bonding free energy to determine the fundamental properties of materials. Applied to carbon nanoclusters and graphynes, our model not only precisely predicts bonding energies and electron density without relying on external parameters, but also enhances the prediction of electronic structures through bond occupancy numbers, which act as effective hopping integrals. This innovation offers insights into the structural properties and quantum behavior of electrons across various dimensions.

cond-mat.mes-hall↗

arXiv2024

Classical-quantum correspondence in the noise-based dissipative systems

We investigate the correspondence between classical noise and quantum environments. Although it has been known that the classical noise can be mapped to the quantum environments only for pure dephasing and infinite-temperature dissipation processes, we describe that this limitation can be circumvented by introducing auxiliary systems and conservation. Taking a two-level system as an example, we construct the so-called central spin model with its couplings fluctuating as the classical noise, and then acquire its statistical-average dynamics which captures the dissipations beyond the infinite temperature. By adjusting the number of the auxiliary systems and their initial states, the noise-based model reproduces both Markovian and non-Markovian evolutions. It is also found that different quantities of the two-level system are governed by different model parameters, indicating that the constructed model is an efficient simulator for specific observables, rather than an equivalent form of a realistic open system. In addition, the model is also applicable to investigate topical mechanisms of the open systems, e.g. negative temperatures and asymmetric equidistant quenches.

quant-ph↗

arXiv2023

Strange memory effect of low-field microwave absorption in copper-substituted lead apatite

We observe a considerable hysteresis effect of low-field microwave absorption (LFMA) in copper-substituted lead apatite. By continuously rotating samples under external magnetic field, this effect is diminished which can not be renewed by a strong magnetic field but will be spontaneously recovered after two days, indicating its glassy features and excluding possibility of any ferromagnetism. The intensity of LFMA is found to sharply decrease at around 250K, suggesting a phase transition takes place. A lattice gauge model is then employed to assign these effects to the transition between superconducting Meissner phase and vortex glass, and the slow dynamics wherein is calculated as well.

cond-mat.supr-con↗

arXiv2023

Spatial inversion symmetry breaking of vortex current in biased-ladder superfluid

We investigate the quench dynamics of interacting bosons on a two-leg ladder in presence of a uniform Abelian gauge field. The model hosts a variety of emergent quantum phases, and we focus on the superfluid biased-ladder phase breaking the $Z_{2}$ symmetry of two legs. We observe an asymmetric spreading of vortex current and particle density, i.e., the current behaves particle-like on the right and wave-like on the left, indicating spontaneous breaking of the spatial inversion symmetry. By decreasing the repulsion strength, it is found the particle-like current is more robust than the wave-like one. The evolution of entanglement entropy manifests logarithmic growth with time suggesting many-body localization matters.

cond-mat.quant-gas↗

arXiv2022

Dynamics of entangled domain walls in quantum many-body scars and breakdown of their prethermalization

Based on the PXP model adapted for Rydberg-blockaded chains, we investigate dynamics of topological domain walls between different quantum many-body scar states of $\mathbb{Z}_2$ symmetry. It is found that, the domain walls not only possess oscillating features of scars but also manifest longstanding bipartite entanglement with exactly unchanged flip-flop phase difference, suggesting their potential as quantum information resource. A periodically driven field is exerted and the high-frequency drive gives rise to a phase transition from prethermalization to Floquet localization. In order to investigate the stability of domain walls acting as information carriers, we further simulate the collision between them and find negligible influence on each other. Subsequently, the quench dynamics with domain walls reveals exotic physics and applicable potentials of nonthermalized scar states.

quant-ph↗

arXiv2021

Critical region of topological trivial and nontrivial phases in interacting Kitaev chain with spatially varying potentials

By using the variational matrix product state method, we numerically study the interacting Kitaev chain with spatially varying periodic and quasi-periodic potentials and the latter follows the Fibonacci sequence. The edge correlation functions of Majorana fermions and low-lying ground states are computed to explore the robustness of topological superconducting phase. It is found that the original topological nontrivial phase is separated into to two branches by an emergent topological trivial phase, as a result of the competition among spatially varying potential, electronic Coulomb interaction and chemical potential. The analysis of energy gap and occupation number together suggests that the spontaneous symmetry breaking and the lift of degeneracy in the topological trivial phase are enabled by a potential-induced Fracton mechanism, namely the pairing of four Majorana fermions. It can be broken by further enhancing the interaction, and then the nontrivial phase reemerges. The evolution from the emergent fractal structure of population outside the critical region to the original structure of charge density wave is investigated as well.

cond-mat.str-el↗

arXiv2019

A Locking-free DP-Q2-P1 MFEM for Incompressible Nonlinear Elasticity Problems

A mixed finite element method (MFEM), using dual-parametric piecewise bi-quadratic and affine (DP-Q2-P1) finite element approximations for the deformation and the pressure like Lagrange multiplier respectively, is developed and analyzed for the numerical computation of incompressible nonlinear elasticity problems with large deformation gradient, and a damped Newton method is applied to solve the resulted discrete problem. The method is proved to be locking free and stable. The accuracy and efficiency of the method are illustrated by numerical experiments on some typical cavitation problems.

math.AP↗

arXiv2018

A Vector-Based Representation of the Chemical Bond for the Substituted Torsion of Biphenyl

We use a new interpretation of the chemical bond within QTAIM. The bond-path framework set $\mathbb{B} = \{p, q, r\}$ with associated linkages with lengths $\mathbb{H}^*,m\mathbb{H}$ and the familiar bond-path length is used to describe a torsion $θ$, $0.0^{\circ} \leq θ\lt 22.0^{\circ}$ of \emph{para}-substituted biphenyl, $\mathrm{C}_{12}\mathrm{H}_{9-x}$, $x = \mathrm{N}(\mathrm{CH}_3)_2$, $\mathrm{NH}_2$, $\mathrm{CH}_3$, CHO, CN, $\mathrm{NO}_{2}$. We include consideration of the H---H bonding interactions and find that the lengths $\mathbb{H} \gt \mathbb{H}^{*}$ that we explain in terms of the most and least preferred directions of charge density accumulation. We also consider the fractional eigenvector-following path lengths $\mathbb{H}_f$ and $\mathbb{H}_{f{θ_{\rm min}}}$.

physics.chem-ph↗

arXiv2018

A Mixed Finite Element Method for Multi-Cavity Computation in Incompressible Nonlinear Elasticity

A mixed finite element method combining an iso-parametric $Q_2$-$P_1$ element and an iso-parametric $P_2^+$-$P_1$ element is developed for the computation of multiple cavities in incompressible nonlinear elasticity. The method is analytically proved to be locking-free and convergent, and it is also shown to be numerically accurate and efficient by numerical experiments. Furthermore, the newly developed accurate method enables us to find an interesting new bifurcation phenomenon in multi-cavity growth.

math.AP↗

arXiv2016

Character-level Convolutional Network for Text Classification Applied to Chinese Corpus

This article provides an interesting exploration of character-level convolutional neural network solving Chinese corpus text classification problem. We constructed a large-scale Chinese language dataset, and the result shows that character-level convolutional neural network works better on Chinese corpus than its corresponding pinyin format dataset. This is the first time that character-level convolutional neural network applied to text classification problem.

cs.CL↗