arXiv subjects

Xiangwen Wang

Publications and source records attributed to Xiangwen Wang.

11 recordsLinked to original sources

arXiv2026

Symmetry-aware generative design of flat-band materials beyond known crystal-net prototypes

Flat electronic bands underlie a range of strongly correlated and topological phenomena, whose design in real materials has so far relied on a small catalogue of named geometric motifs such as kagome, Lieb, and pyrochlore nets. This discrete catalogue is by no means to exhaust the geometries that support flat bands in real compounds, as band flatness is a property of network connectivity. Here we combine a continuous geometric representation of crystal sublattices, with a symmetry-constrained generative model, to access a broader design space for materials hosting flat bands. The key step is to choose sublattice motifs that are outside the known geometric clusters, ensuring the novelty of the generated structures. We then introduce SkeleGen, which pins these unconventional skeletons to symmetry-compatible Wyckoff positions while denoising the surrounding chemistry, resulting in 9,352 crystal candidates that survive stability and flatnessscreening. Band flatness is confirmed using high throughput full DFT calculations, which agree well also with the tight-binding spectra of the isolated skeletons, supporting a geometric origin of the band flatness. We demonstrate "out-of-distribution" motifs as a new design principle to dramatically expand geometric repertoire for materials discovery, potentially beyond flat bands.

cond-mat.mtrl-sci↗

arXiv2026

Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models

Large language models remain vulnerable to jailbreak attacks, yet we still lack a systematic understanding of how jailbreak success scales with attacker effort across methods, model families, and harm types. We initiate a scaling-law framework for jailbreaks by treating each attack as a compute-bounded optimization procedure and measuring progress on a shared FLOPs axis. Our systematic evaluation spans four representative jailbreak paradigms, covering optimization-based attacks, self-refinement prompting, sampling-based selection, and genetic optimization, across multiple model families and scales on a diverse set of harmful goals. We investigate scaling laws that relate attacker budget to attack success score by fitting a simple saturating exponential function to FLOPs--success trajectories, and we derive comparable efficiency summaries from the fitted curves. Empirically, prompting-based paradigms tend to be the most compute-efficient compared to optimization-based methods. To explain this gap, we cast prompt-based updates into an optimization view and show via a same-state comparison that prompt-based attacks more effectively optimize in prompt space. We also show that attacks occupy distinct success--stealthiness operating points with prompting-based methods occupying the high-success, high-stealth region. Finally, we find that vulnerability is strongly goal-dependent: harms involving misinformation are typically easier to elicit than other non-misinformation harms.

cs.LG↗

arXiv2026

Aligning Compound AI Systems via System-level DPO

Compound AI systems, comprising multiple interacting components such as LLMs, foundation models, and external tools, have demonstrated remarkable improvements compared to single models in various tasks. To ensure their effective deployment in real-world applications, aligning these systems with human preferences is crucial. However, aligning the compound system via policy optimization, unlike the alignment of a single model, is challenging for two main reasons: (i) non-differentiable interactions between components make end-to-end gradient-based optimization method inapplicable, and (ii) system-level preferences cannot be directly transformed into component-level preferences. To address these challenges, we first formulate compound AI systems as Directed Acyclic Graphs (DAGs), explicitly modeling both component interactions and the associated data flows. Building on this formulation, we introduce $\textbf{SysDPO}$, a framework that extends Direct Preference Optimization (DPO) to enable joint system-level alignment. We propose two variants, SysDPO-Direct and SysDPO-Sampling, tailored for scenarios depending on whether we construct a system-specific preference dataset. We empirically demonstrate the effectiveness of our approach across two applications: the joint alignment of a language model and a diffusion model, and the joint alignment of an LLM collaboration system.

cs.LG↗

arXiv2025

ThermoRL:Structure-Aware Reinforcement Learning for Protein Mutation Design to Enhance Thermostability

Designing mutations to optimize protein thermostability remains challenging due to the complex relationship between sequence variations, structural dynamics, and thermostability, often assessed by δδG (the change in free energy of unfolding). Existing methods rely on experimental random mutagenesis or prediction models tested with pre-defined datasets, using sequence-based heuristics and treating enzyme design as a one-step process without iterative refinement, which limits design space exploration and restricts discoveries beyond known variations. We present ThermoRL, a framework based on reinforcement learning (RL) that leverages graph neural networks (GNN) to design mutations with enhanced thermostability. It combines a pre-trained GNN-based encoder with a hierarchical Q-learning network and employs a surrogate model for reward feedback, guiding the RL agent on where (the position) and which (mutant amino acid) to apply for enhanced thermostability. Experimental results show that ThermoRL achieves higher or comparable rewards than baselines while maintaining computational efficiency. It filters out destabilizing mutations and identifies stabilizing mutations aligned with experimental data. Moreover, ThermoRL accurately detects key mutation sites in unseen proteins, highlighting its strong generalizability. This RL-guided approach powered by GNN embeddings offers a robust alternative to traditional protein mutation design.

cs.CE↗

arXiv2025

Structure-Informed Learning of Flat Band 2D Materials

Flat electronic bands enhance electron-electron interactions and give rise to correlated states such as unconventional superconductivity or fractional topological phases. However, most current efforts towards flat-band materials discovery rely on density functional theory (DFT) calculations and manual band structures inspection, restraining their applicability to vast unexplored material spaces. While data-driven methods offer a scalable alternative, most existing models either depend on band structure inputs or focus on scalar properties like bandgap, which fail to capture flat-band characteristics. Here, we report a structure-informed framework for the discovery of previously unrecognized flat-band two-dimensional (2D) materials, which combines a data-driven flatness score capturing both band dispersion and density-of-states characteristics with multi-modal learning from atomic structure inputs. The framework successfully identified multiple flat-band candidates, with DFT validation of kagome-based systems confirming both band flatness and topological character. Our results show that the flatness score provides a physically meaningful signal for identifying flat bands from atomic geometry. The framework uncovers multiple new candidates with topologically nontrivial flat bands from unlabeled data, with consistent model performance across structurally diverse materials. By eliminating the need for precomputed electronic structures, our method enables large-scale screening of flat-band materials and expands the search space for discovering strongly correlated quantum materials.

cond-mat.mtrl-sci↗

arXiv2021

Bridging atomistic simulations and thermodynamic hydration models of aqueous electrolyte solutions

Chemical thermodynamic models of solvent and solute activities predict the equilibrium behaviour of aqueous solutions. How-ever, these models are semi-empirical. They represent micro-scale ion and solvent behaviours that control the macroscopic properties using small numbers of parameters whose values are obtained by fitting to activities and other partial derivatives of the Gibbs energy measured for the bulk solutions. We have conducted atomistic simulations of aqueous electrolyte solutions (MgCl2 and CaCl2) to determine the parameters of aqueous thermodynamic hydration models. We have implemented a coopera-tive hydration model to categorize the water molecules in electrolyte solutions into different subpopulations. The value of the electrolyte-specific parameter, k, was determined from the ion-affected subpopulation with the lowest absolute value of the free energy of removing the water molecule. The other equilibrium constant parameter, K1, associated with the first degree of hydra-tion, was computed from the free energy of hydration of hydrated clusters in solution. The hydration number, h, was determined from a reorientation dynamic analysis of the water subpopulations compared to bulk-like behaviour. The computed values of these parameters were inserted in the Stokes & Robinson (Journal of Solution Chemistry 1973, 2, 173-191) and Balomenos (Fluid Phase Equilibria 2006, 243, 29-37) models, which were applied to evaluate the osmotic coefficients of MgCl2 solutions. Such an approach removes the dependence on the availability of experimental data and could lead to aqueous thermodynamic models capable of estimating the values of solute and solvent activities, thermal and volumetric properties for a wide range of composition and concentrations.

physics.chem-ph↗

arXiv2019

Online Gambling of Pure Chance: Wager Distribution, Risk Attitude, and Anomalous Diffusion

Online gambling sites offer many different gambling games. In this work we analyse the gambling logs of numerous solely probability-based gambling games and extract the wager and odds distributions. We find that the log-normal distribution describes the wager distribution at the aggregate level. Viewing the gamblers' net incomes as random walks, we study the mean-squared displacement of net income and related quantities and find different diffusive behaviors for different games. We discuss possible origins for the observed anomalous diffusion.

physics.soc-ph↗

arXiv2018

Behavior analysis of virtual item gambling

From the gambling logs of an online lottery game we extract the probability distribution of various quantities (e.g., bet value, total pool size, waiting time between successive gambles) as well as related correlation coefficients. We view the net change of income of each player as a random walk. The mean squared displacement of these net income random walks exhibits a transition between a super-diffusive and a normal diffusive regime. We discuss different random walk models with truncated power-law step lengths distributions that allow to reproduce some of the properties extracted from the gambling logs. Analyzing the mean squared displacement and the first-passage time distribution for these models allows to identify the key features needed for observing this crossover from super-diffusion to normal diffusion.

physics.soc-ph↗

arXiv2017

Foraging patterns in online searches

Nowadays online searches are undeniably the most common form of information gathering, as witnessed by billions of clicks generated each day on search engines. In this work we describe online searches as foraging processes that take place on the semi-infinite line. Using a variety of quantities like probability distributions and complementary cumulative distribution functions of step-length and waiting time as well as mean square displacements and entropies, we analyze three different click-through logs that contain the detailed information of millions of queries submitted to search engines. Notable differences between the different logs reveal an increased efficiency of the search engines. In the language of foraging, the newer logs indicate that online searches overwhelmingly yield local searches (i.e. on one page of links provided by the search engines), whereas for the older logs the foraging processes are a combination of local searches and relocation phases that are power law distributed. Our investigation of click logs of search engines therefore highlights the presence of intermittent search processes (where phases of local explorations are separated by power law distributed relocation jumps) in online searches. It follows that good search engines enable the users to find the information they are looking for through a local exploration of a single page with search results, whereas for poor search engines users are often forced to do a broader exploration of different pages.

physics.soc-ph↗

arXiv2015

Relaxation processes in a system with logarithmic growth

We discuss relaxation and aging processes in the one- and two-dimensional $ABC$ models. In these driven diffusive systems of three particle types, biased exchanges in one direction yield a coarsening process characterized in the long time limit by a logarithmic growth of ordered domains that take the form of stripes. From the time-dependent length, derived from the equal-time spatial correlator, and from the mean displacement of individual particles different regimes in the formation and growth of these domains can be identified. Analysis of two-times correlation and response functions reveals dynamical scaling in the asymptotic logarithmic growth regime as well as complicated finite-time and finite-size effects in the early and intermediate time regimes.

cond-mat.stat-mech↗

arXiv2013

Growth of scale-free networks under heterogeneous control

Real-life networks often encounter vertex dysfunctions, which are usually followed by recoveries after appropriate maintenances. In this paper we present our research on a model of scale-free networks whose vertices are regularly removed and put back. Both the frequency and length of time of the disappearance of each vertex depend on the degree of the vertex, creating a heterogeneous control over the network. Our simulation results show very interesting growth pattern of this kind of networks. We also find that the scale-free property of the degree distribution is maintained in the proposed heterogeneously controlled networks. However, the overall growth rate of the networks in our model can be remarkably reduced if the inactive periods of the vertices are kept long.

physics.soc-ph↗