arXiv subjects

Yin Song

Publications and source records attributed to Yin Song.

15 recordsLinked to original sources

arXiv2025

STED and Consistency Scoring: A Framework for Evaluating LLM Structured Output Reliability

Large Language Models (LLMs) are increasingly deployed for structured data generation, yet output consistency remains critical for production applications. We introduce a comprehensive framework for evaluating and improving consistency in LLM-generated structured outputs. Our approach combines: (1) STED (Semantic Tree Edit Distance), a novel similarity metric balancing semantic flexibility with structural strictness when comparing JSON outputs, and (2) a consistency scoring framework aggregating multiple STED measurements across repeated generations to quantify reliability. Through systematic experiments on synthetic datasets with controlled schema, expression, and semantic variations, we demonstrate STED achieves superior performance ($0.86-0.90$ similarity for semantic equivalents, $0.0$ for structural breaks) compared to existing metrics including TED, BERTScore, and DeepDiff. Applying our framework to benchmark six LLMs reveals significant variations: Claude-3.7-Sonnet demonstrates exceptional consistency, maintaining near-perfect structural reliability even at high temperatures ($T=0.9$), while models like Claude-3-Haiku and Nova-Pro exhibit substantial degradation requiring careful tuning. Our framework enables practical applications including targeted model selection for structured tasks, iterative prompt refinement for reproducible results, and diagnostic analysis to identify inconsistency root causes. This work provides theoretical foundations and practical tools for ensuring reliable structured output generation in LLM-based production systems.

cs.CL↗

arXiv2025

HoSGFEM: High-order stable generalized finite element method for elliptic interface problem

The Generalized Finite Element Method (GFEM) is an effective unfitted numerical method for handling interface problems. By augmenting the standard FEM space with an appropriate enrichment space, GFEM can accurately capture C^0 solutions across the interfaces. While numerous GFEMs for interface problems have been studied, establishing a stable high-order GFEM with optimal convergence rates and robust system conditioning remains a challenge. The highest known order of two was established by Zhang and Babu\v{s}ka (SGFEM2, Comput. Methods Appl. Mech. Engrg. 363 (2020), 112889). In this paper, we propose a unified enrichment space construction and establish arbitrary high-order stable GFEMs (HoSGFEM) for elliptic interface problems. The main idea distinguishes itself from Zhang and Babu\v{s}ka's SGFEM2 substantially and it is twofold: a) we construct dimensionality-reduced auxiliary locally supported piecewise polynomials that satisfy the partition of unity property for elements containing interfaces; b) we construct the enrichment scheme based on d{1,(x-x_c^e),...,(y-y_c^e)^{p-1}} (d is the distance function; (x_c^e, y_c^e) is the center of the element containing interface, thus element-based) for arbitrary p-th order elements instead of d, d{1,x,y} or d{1,x,y,x^2,xy,y^2} (global functions) for p=1,2 in the literature. This idea results in an enrichment space that has a large angle with the standard FEM space, leading to the stability of the method with system condition number growing in order O(h^{-2}). We establish optimal convergence rates for HoSGFEM solutions under the proposed construction. Various numerical experiments with both straight and curved interfaces demonstrate the optimal convergence, FEM-comparable system condition number with O(h^{-2}) growth, and robustness as element boundaries approach interfaces.

math.NA↗

arXiv2025

Scaling Context, Not Parameters: Training a Compact 7B Language Model for Efficient Long-Context Processing

We present MegaBeam-Mistral-7B, a language model that supports 512K-token context length. Our work addresses practical limitations in long-context training, supporting real-world tasks such as compliance monitoring and verification. Evaluated on three long-context benchmarks, our 7B-parameter model demonstrates superior in-context learning performance on HELMET and robust retrieval and tracing capability on RULER. It is currently the only open model to achieve competitive long-range reasoning on BABILong at 512K context length without RAG or targeted fine-tuning. Released as fully open source under the Apache 2.0 license, the model has been downloaded over 100,000 times on Hugging Face. Model available at: https://huggingface.co/aws-prototyping/MegaBeam-Mistral-7B-512k

cs.CL↗

arXiv2025

Stable quadratic generalized IsoGeometric analysis for elliptic interface problem

Unfitted mesh formulations for interface problems generally adopt two distinct methodologies: (i) penalty-based approaches and (ii) explicit enrichment space techniques. While Stable Generalized Finite Element Method (SGFEM) has been rigorously established for one-dimensional and linear-element cases, the construction of optimal enrichment spaces preserving approximation-theoretic properties within isogeometric analysis (IGA) frameworks remains an open challenge. In this paper, we introduce a stable quadratic generalized isogeometric analysis (SGIGA2) for two-dimensional elliptic interface problems. The method is achieved through two key ideas: a new quasi-interpolation for the function with C0 continuous along interface and a new enrichment space with controlled condition number for the stiffness matrix. We mathematically prove that the present method has optimal convergence rates for elliptic interface problems and demonstrate its stability and robustness through numerical verification.

math.NA↗

arXiv2024

Anomalously reduced homogeneous broadening of two-dimensional electronic spectroscopy at high temperature by detailed balance

Dissipation and decoherence of quantum systems in thermal environments is important to various spectroscopies. It is generally believed that dissipation can broaden the line shape of spectroscopies, and thus stronger system-bath interaction can result in more significant homogeneous broadening of two-dimensional electronic spectroscopy (2DES). Here we show that the case can be the opposite in the regime of electromagnetically induced transparency (EIT). We predict that assisted by EIT, the homogeneous broadening of the 2DES at a higher temperature can be significantly reduced due to the detailed balance. This anomalous effect is due to the long-lasting off-diagonal peaks in 2DES.

quant-ph↗

arXiv2022

Charge Separation in the Photosystem II Reaction Center Resolved by Multispectral Two-Dimensional Electronic Spectroscopy

The photosystem II reaction center (PSII-RC) performs the primary energy conversion steps of oxygenic photosynthesis. While the PSII-RC has been studied extensively, the similar timescales of energy transfer and charge separation, and the severely overlapping pigment transitions in the Qy region have led to multiple models of its charge separation mechanism and excitonic structure. Here we combine two-dimensional electronic spectroscopy (2DES) with a continuum probe and two-dimensional electronic vibrational spectroscopy (2DEV) to study the cyt b559-D1D2 PSII-RC at 77K. This multispectral combination correlates the overlapping Qy excitons with distinct anion and pigment-specific Qx and mid-IR transitions to resolve the charge separation mechanism and excitonic structure. Through extensive simultaneous analysis of the multispectral 2D data we find that charge separation proceeds on multiple timescales from a highly delocalized excited state via a single pathway in which PheoD1 is the primary electron acceptor, while ChlD1 and PD1 act in concert as the primary electron donor.

physics.chem-ph↗

arXiv2022

Learning from Drivers to Tackle the Amazon Last Mile Routing Research Challenge

The goal of the Amazon Last Mile Routing Research Challenge is to integrate the real-life experience of Amazon drivers into the solution of optimal route planning and optimization. This paper presents our method that tackles this challenge by hierarchically combining machine learning and conventional Traveling Salesperson Problem (TSP) solvers. Our method reaps the benefits from both worlds. On the one hand, our method encodes driver know-how by learning a sequential probability model from historical routes at the zone level, where each zone contains a few parcel stops. It then uses a single step policy iteration method, known as the Rollout algorithm, to generate plausible zone sequences sampled from the learned probability model. On the other hand, our method utilizes proven methods developed in the rich TSP literature to sequence stops within each zone efficiently. The outcome of such a combination appeared to be promising. Our method obtained an evaluation score of $0.0374$, which is comparable to what the top three teams have achieved on the official Challenge leaderboard. Moreover, our learning-based method is applicable to driving routes that may exhibit distinct sequential patterns beyond the scope of this Challenge. The source code of our method is publicly available at https://github.com/aws-samples/amazon-sagemaker-amazon-routing-challenge-sol

cs.AI↗

arXiv2021

Isolating Polaritonic 2D-IR Transmission Spectra

Strong coupling between vibrational transitions in molecules within a resonant optical microcavity leads to the formation of collective, delocalized vibrational polaritons. There are many potential applications of "polaritonic chemistry," ranging from modified chemical reactivity to quantum in-formation processing. One challenge in obtaining the polaritonic response is to remove a background contribution due to the uncoupled molecules that generate an ordinary 2D-IR spectrum whose amplitude is filtered by the polariton transmission spectrum. We show that most features in 2D-IR spectra of vibrational polaritons can be explained by a linear superposition of this background signal and the true polariton response. Through a straightforward correction procedure, where the filtered bare molecule 2D-IR spectrum is subtracted from the measured cavity response, we recover the polaritonic spectrum.

physics.chem-ph↗

arXiv2020

Spectral Dynamic Causal Modelling of Resting-State fMRI: Relating Effective Brain Connectivity in the Default Mode Network to Genetics

We conduct an imaging genetics study to explore how effective brain connectivity in the default mode network (DMN) may be related to genetics within the context of Alzheimer's disease and mild cognitive impairment. We develop an analysis of longitudinal resting-state functional magnetic resonance imaging (rs-fMRI) and genetic data obtained from a sample of 111 subjects with a total of 319 rs-fMRI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. A Dynamic Causal Model (DCM) is fit to the rs-fMRI scans to estimate effective brain connectivity within the DMN and related to a set of single nucleotide polymorphisms (SNPs) contained in an empirical disease-constrained set which is obtained out-of-sample from 663 ADNI subjects having only genome-wide data. We examine longitudinal data in both a 4-region and an 6-region network and relate longitudinal effective brain connectivity networks estimated using spectral DCM to SNPs using both linear mixed effect (LME) models as well as function-on-scalar regression (FSR). In the former case we implement a parametric bootstrap for testing SNP coefficients and make comparisons with p-values obtained from the chi-squared null distribution. We also implement a parametric bootstrap approach for testing regression functions in FSR and we make comparisons between p-values obtained from the parametric bootstrap to p-values obtained using the F-distribution with degrees-of-freedom based on Satterthwaite's approximation. In both networks we report on exploratory patterns of associations with relatively high ranks that exhibit stability to the differing assumptions made by both FSR and LME.

q-bio.NC↗

arXiv2020

A Bayesian Spatial Model for Imaging Genetics

We develop a Bayesian bivariate spatial model for multivariate regression analysis applicable to studies examining the influence of genetic variation on brain structure. Our model is motivated by an imaging genetics study of the Alzheimer's Disease Neuroimaging Initiative (ADNI), where the objective is to examine the association between images of volumetric and cortical thickness values summarizing the structure of the brain as measured by magnetic resonance imaging (MRI) and a set of 486 SNPs from 33 Alzheimer's Disease (AD) candidate genes obtained from 632 subjects. A bivariate spatial process model is developed to accommodate the correlation structures typically seen in structural brain imaging data. First, we allow for spatial correlation on a graph structure in the imaging phenotypes obtained from a neighbourhood matrix for measures on the same hemisphere of the brain. Second, we allow for correlation in the same measures obtained from different hemispheres (left/right) of the brain. We develop a mean-field variational Bayes algorithm and a Gibbs sampling algorithm to fit the model. We also incorporate Bayesian false discovery rate (FDR) procedures to select SNPs. We implement the methodology in a new release of the R package bgsmtr. We show that the new spatial model demonstrates superior performance over a standard model in our application. Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu).

stat.ME↗

arXiv2019

A Potts-Mixture Spatiotemporal Joint Model for Combined MEG and EEG Data

We develop a new methodology for determining the location and dynamics of brain activity from combined magnetoencephalography (MEG) and electroencephalography (EEG) data. The resulting inverse problem is ill-posed and is one of the most difficult problems in neuroimaging data analysis. In our development we propose a solution that combines the data from three different modalities, MRI, MEG, and EEG, together. We propose a new Bayesian spatial finite mixture model that builds on the mesostate-space model developed by Daunizeau and Friston (2007). Our new model incorporates two major extensions: (i) We combine EEG and MEG data together and formulate a joint model for dealing with the two modalities simultaneously; (ii) we incorporate the Potts model to represent the spatial dependence in an allocation process that partitions the cortical surface into a small number of latent states termed mesostates. The cortical surface is obtained from MRI. We formulate the new spatiotemporal model and derive an efficient procedure for simultaneous point estimation and model selection based on the iterated conditional modes algorithm combined with local polynomial smoothing. The proposed method results in a novel estimator for the number of mixture components and is able to select active brain regions which correspond to active variables in a high-dimensional dynamic linear model. The methodology is investigated using synthetic data and simulation studies and then demonstrated on an application examining the neural response to the perception of scrambled faces. R software implementing the methodology along with several sample datasets are available at the following GitHub repository https://github.com/v2south/PottsMix.

stat.AP↗

arXiv2017

Bipartite entanglement in AJL's algorithm for three-strand braids

Aharonov, Jones, and Landau [Algorithmica 55, 395 (2009)] have presented a polynomial quantum algorithm for approximating the Jones polynomial. We investigate the bipartite entanglement properties in AJL's algorithm for three-strand braids. We re-describe AJL's algorithm as an equivalent algorithm which involves three work qubits in some mixed state coupled to a single control qubit. Furthermore, we use the Peres entanglement criterion to study the entanglement features of the state before measurements present in the re-described algorithm for all bipartitions. We show that the state is a product state relative to the bipartition between the first work qubit and the others. And it has no entanglement between the control qubit and work ones. We also prove a sufficient and necessary condition for its entanglement between the second (third) work qubit and the others. Moreover, we discuss the relation between its bipartite entanglement and elementary crossings in the three-strand braid group. We find that braids whose trace closures are topologically identical may have different entanglement properties in AJL's algorithm.

quant-ph↗

arXiv2014

Effects of swift heavy ions irradiation parameters on optical properties of muscovite mica

Muscovite mica sheets with a thickness of 25 μm were irradiated by various kinds of swift heavy ions (Sn, Xe and Bi) in HIRFL. The fluences ranged from 1$\times$10^{10} ions/cm^2 to 8$\times$10^{11} ions/cm^2. The electronic energy loss (dE/dx)_e was increased from 14.7 keV/nm to 31.2 keV/nm. The band gap and Urbach energy of pristine and irradiated mica were analyzed by ultraviolet- visible spectroscopy. Periodic fringes in long wave length of the absorption spectra caused by interference phenomenon, were disturbed as the (dE/dx)_e increased. It was suggested that the chemical bonds between Tetrahedral-Octohedral-Tetrahedral (TOT) layers of mica were destroyed. Thus the smooth surface was cleaved after irradiation. The band gap was narrowed down with the increasing (dE/dx)_e and fluences. The values of Urbach energy were increased as the (dE/dx)_e and fluences gradually increased. It was indicated that the amount of defects and the proportion of amorphous structure were increased in mica irradiated under increased (dE/dx)_e and fluences. Fluences took a distinctly important role in optical properties of mica.

cond-mat.mtrl-sci↗

arXiv2013

Learning Hidden Structures with Relational Models by Adequately Involving Rich Information in A Network

Effectively modelling hidden structures in a network is very practical but theoretically challenging. Existing relational models only involve very limited information, namely the binary directional link data, embedded in a network to learn hidden networking structures. There is other rich and meaningful information (e.g., various attributes of entities and more granular information than binary elements such as "like" or "dislike") missed, which play a critical role in forming and understanding relations in a network. In this work, we propose an informative relational model (InfRM) framework to adequately involve rich information and its granularity in a network, including metadata information about each entity and various forms of link data. Firstly, an effective metadata information incorporation method is employed on the prior information from relational models MMSB and LFRM. This is to encourage the entities with similar metadata information to have similar hidden structures. Secondly, we propose various solutions to cater for alternative forms of link data. Substantial efforts have been made towards modelling appropriateness and efficiency, for example, using conjugate priors. We evaluate our framework and its inference algorithms in different datasets, which shows the generality and effectiveness of our models in capturing implicit structures in networks.

cs.LG↗

arXiv2013

Characterizing A Database of Sequential Behaviors with Latent Dirichlet Hidden Markov Models

This paper proposes a generative model, the latent Dirichlet hidden Markov models (LDHMM), for characterizing a database of sequential behaviors (sequences). LDHMMs posit that each sequence is generated by an underlying Markov chain process, which are controlled by the corresponding parameters (i.e., the initial state vector, transition matrix and the emission matrix). These sequence-level latent parameters for each sequence are modeled as latent Dirichlet random variables and parameterized by a set of deterministic database-level hyper-parameters. Through this way, we expect to model the sequence in two levels: the database level by deterministic hyper-parameters and the sequence-level by latent parameters. To learn the deterministic hyper-parameters and approximate posteriors of parameters in LDHMMs, we propose an iterative algorithm under the variational EM framework, which consists of E and M steps. We examine two different schemes, the fully-factorized and partially-factorized forms, for the framework, based on different assumptions. We present empirical results of behavior modeling and sequence classification on three real-world data sets, and compare them to other related models. The experimental results prove that the proposed LDHMMs produce better generalization performance in terms of log-likelihood and deliver competitive results on the sequence classification problem.

stat.ML↗