arXiv subjects

Shuyang Cao

Publications and source records attributed to Shuyang Cao.

At least 19 recordsLinked to original sources

arXiv2026

Fast Computation of Nested Cross-Validation for Penalized Regression

Cross-validation is a resampling procedure that provides a point estimate of generalization error for any predictive model. Cross-validation is widely used for model selection and evaluation. Uncertainty in the cross-validation estimate is challenging to quantify, and estimation of its variance is known to require multiple runs of the entire resampling procedure. Nested cross-validation computes a prediction interval for the generalization error for a given model and training dataset by resampling the entire cross-validation procedure but incurs extraordinary computational cost. We provide an efficient method for computing the nested cross-validation prediction interval using only a single model fit for some penalized regression models including ridge regression, spline smoothing, and some functional regression models. We characterize when our proposed method should be expected to out-perform resampling-based nested cross-validation in various scaling regimes as well as in finite samples. Experiments for functional principle components regression demonstrate non-trivial cases in which our proposed method improves run times substantially.

stat.CO↗

arXiv2025

ExpertLongBench: Benchmarking Language Models on Expert-Level Long-Form Generation Tasks with Structured Checklists

This paper introduces ExpertLongBench, an expert-level benchmark containing 11 tasks from 9 domains that reflect realistic expert workflows and applications. Beyond question answering, the application-driven tasks in ExpertLongBench demand long-form outputs that can exceed 5,000 tokens and strict adherence to domain-specific requirements. Notably, each task in ExpertLongBench includes a rubric, designed or validated by domain experts, to specify task requirements and guide output evaluation. Furthermore, we propose CLEAR, an evaluation framework that supports accurate evaluation of long-form model outputs in our benchmark. To achieve fine-grained, expert-aligned evaluation, CLEAR derives checklists from both model outputs and references by extracting information corresponding to items in the task-specific rubric. Checklist items of model outputs are then compared with corresponding items of reference outputs to assess their correctness, enabling grounded evaluation. We benchmark 13 popular large language models (LLMs) and analyze components in CLEAR, showing that (1) existing LLMs, with the top performer Gemini-2.5-Pro achieving only a 33.4 F1 score, require significant improvement for expert-level tasks; (2) models can generate content corresponding to the required aspects, but far from correct; and (3) accurate checklist extraction and comparison in CLEAR can be achieved by open-weight models for more scalable, reproducible, and low-cost usage.

cs.CL↗

arXiv2025

Field mixing in a thermal medium: A quantum master equation approach

We studied the nonequilibrium dynamics of the indirect mixing of two (pseudo-)scalar fields induced by their couplings to common decay channels in a medium. The effective non-Markovian quantum master equation (QME) for the two fields' reduced density matrix is derived to leading order in the couplings of the two fields with the medium, but to all orders of the couplings among degrees of freedom in the medium. The self-energy and noise-kernel in the QME satisfy a fluctuation-dissipation relation. The solutions show that an initial expectation value (condensate) of one field induces a condensate of the other field through the indirect mixing and that the populations and coherence of the two fields thermalize and approach to non-vanishing values asymptotically. The nearly-degenerate field masses and coupling strengths resonantly enhance the quantum beats and asymptotic coherence, and induce a prominent dynamics of the vacuum after the switch-on of the couplings. We argue that a time-dependent definitions of particles due to the changing vacuum must be introduced so as to obtain results consistent with the calculations of equilibrium states in the asymptotic limit. A coupling strength hierarchy breaks down the resonant enhancement in the nearly-degenerate case but leads to different power countings of the coupling strengths in the magnitudes of the observables and time-scales in the evolution, suggesting the possibility of detecting extremely long-lived particles using prepared short-lived particles within a practical experimental period.

quant-ph↗

arXiv2025

Evaluating the Retrieval Robustness of Large Language Models

Retrieval-augmented generation (RAG) generally enhances large language models' (LLMs) ability to solve knowledge-intensive tasks. But RAG may also lead to performance degradation due to imperfect retrieval and the model's limited ability to leverage retrieved content. In this work, we evaluate the robustness of LLMs in practical RAG setups (henceforth retrieval robustness). We focus on three research questions: (1) whether RAG is always better than non-RAG; (2) whether more retrieved documents always lead to better performance; (3) and whether document orders impact results. To facilitate this study, we establish a benchmark of 1500 open-domain questions, each with retrieved documents from Wikipedia. We introduce three robustness metrics, each corresponds to one research question. Our comprehensive experiments, involving 11 LLMs and 3 prompting strategies, reveal that all of these LLMs exhibit surprisingly high retrieval robustness; nonetheless, different degrees of imperfect robustness hinders them from fully utilizing the benefits of RAG.

cs.CL↗

arXiv2025

Condensate decay in a radiation dominated cosmology

We study the decay of a homogeneous condensate of a massive scalar field of mass \emph{m} into massless fields in thermal equilibrium in a radiation dominated cosmology. The model is a \emph{proxy} for the non-equilibrium dynamics of a misaligned axion condensate decaying into radiation. After consistent field quantization in the cosmological background, we obtain the causal equations of motion for a homogeneous condensate including the finite temperature self-energy corrections up to one loop. The dynamical renormalization group is implemented to obtain the time dependent relaxation rate that describes the decay dynamics of the condensate amplitude from stimulated emission and recombination of massless quanta in the medium. It is explicitly shown that a simple friction term in the equation of motion does not describe correctly the decay of the condensate. During the super-Hubble regime, relevant for ultralight dark matter, the condensate amplitude decays as $e^{-\frac{g^2}{10} t^2\,\ln(1/mt)}$. In the sub-Hubble regime the amplitude decays as $e^{-γ(t;T(t))/2}$ with $T(t)=T_0/a(t)$, therefore the finite temperature contribution to the decay rate vanishes fast during the expansion. A main conclusion is that a phenomenological friction term is inadequate to describe the decay in the super-Hubble regime, the decay function $γ(t)$ is always \emph{smaller} than that from a local friction term as a consequence of the cosmological expansion. For ultra light dark matter, the time scale during which transient dynamics is sensitive to cosmological expansion and a local friction term is inadequate, is much longer. A friction term always \emph{underestimates} the time scale of decay in the sub-Hubble case.

hep-th↗

arXiv2025

Is the finite temperature effective potential, effective for dynamics?

We study the applicability of the finite temperature effective potential in the equation of motion of a homogeneous "misaligned" scalar condensate $φ$, and find important caveats that severely restrict its domain of validity: i:) the assumption of local thermodynamic equilibrium (LTE) is in general not warranted, ii:) we show a direct relation between the effective potential and the thermodynamic entropy density $\mathcal{S}= - \partial V_{eff}(T,φ)/\partial T$, which entails that for a dynamical $φ(t)$ the entropy becomes a non-monotonic function of time, iii:) parametric instabilities in both cases with and without spontaneous symmetry breaking lead to profuse particle production with non-thermal distribution functions, iv:) in the case of spontaneous symmetry breaking spinodal instabilities yield a complex effective potential, internal energy and entropy, an untenable situation in thermodynamics. All these caveats associated with dynamical aspects, cannot be overcome by finite temperature equilibrium resummation schemes. We argue that the dynamics of the condensate leads to decoupling and freeze-out from (LTE), and propose a closed quantum system approach based on unitary time evolution. It yields the correct equations of motion without the caveats of the effective potential, and provides a fully renormalized and thermodynamically consistent framework to study the dynamics of the "misaligned" condensate, with real and conserved energy and entropy amenable to numerical study. The evolution of the condensate leads to profuse stimulated particle production with non-thermal distribution functions. Possible emergent asymptotic non-thermal states and eventual re-thermalization are conjectured.

hep-ph↗

arXiv2024

Verifiable Generation with Subsentence-Level Fine-Grained Citations

Verifiable generation requires large language models (LLMs) to cite source documents supporting their outputs, thereby improve output transparency and trustworthiness. Yet, previous work mainly targets the generation of sentence-level citations, lacking specificity about which parts of a sentence are backed by the cited sources. This work studies verifiable generation with subsentence-level fine-grained citations for more precise location of generated content supported by the cited sources. We first present a dataset, SCiFi, comprising 10K Wikipedia paragraphs with subsentence-level citations. Each paragraph is paired with a set of candidate source documents for citation and a query that triggers the generation of the paragraph content. On SCiFi, we evaluate the performance of state-of-the-art LLMs and strategies for processing long documents designed for these models. Our experiment results reveals key factors that could enhance the quality of citations, including the expansion of the source documents' context accessible to the models and the implementation of specialized model tuning.

cs.CL↗

arXiv2024

Is the effective potential, effective for dynamics?

We critically examine the applicability of the effective potential within dynamical situations and find, in short, that the answer is negative. An important caveat of the use of an effective potential in dynamical equations of motion is an explicit violation of energy conservation. An \emph{adiabatic} effective potential is introduced in a consistent quasi-static approximation, and its narrow regime of validity is discussed. Two ubiquitous instances in which even the adiabatic effective potential is not valid in dynamics are studied in detail: parametric amplification in the case of oscillating mean fields, and spinodal instabilities associated with spontaneous symmetry breaking. In both cases profuse particle production is directly linked to the failure of the effective potential to describe the dynamics. We introduce a consistent, renormalized, energy conserving dynamical framework that is amenable to numerical implementation. Energy conservation leads to the emergence of asymptotic highly excited, entangled stationary states from the dynamical evolution. As a corollary, decoherence via dephasing of the density matrix in the adiabatic basis is argued to lead to an emergent entropy, formally equivalent to the entanglement entropy. The results suggest novel characterization of asymptotic equilibrium states in terms of order parameter vs. energy density.

hep-ph↗

arXiv2024

Effective field theory of particle mixing

We introduce an effective field theory to study \emph{indirect} mixing of two fields induced by their couplings to a common decay channel in a medium. The extension of the method of Lee, Oehme and Yang, the cornerstone of analysis of CP violation in flavored mesons, to include mixing of particles with different masses provides a guide to and benchmark for the effective field theory. The analysis reveals subtle caveats in the description of mixing in terms of the widely used non-Hermitian effective Hamiltonian, more acute in the non-degenerate case. The effective field theory describes the dynamics of field mixing where the common intermediate states populate a bath in thermal equilibrium, as an \emph{open quantum system}. We obtain the effective action up to second order in the couplings, where indirect mixing is a consequence of off-diagonal self-energy components. We find that if only one of the mixing fields features an initial expectation value, indirect mixing induces an expectation value of the other field. The equal time two point correlation functions exhibit asymptotic approach to a stationary thermal state, and the emergence of long-lived \emph{bath induced} coherence which display quantum beats as a consequence of interference of quasinormal modes in the medium. The amplitudes of the quantum beats are resonantly enhanced in the nearly degenerate case with potential observational consequences.

hep-ph↗

arXiv2023

AWESOME: GPU Memory-constrained Long Document Summarization using Memory Mechanism and Global Salient Content

Long document summarization systems are critical for domains with lengthy and jargonladen text, yet they present significant challenges to researchers and developers with limited computing resources. Existing solutions mainly focus on efficient attentions or divide-and-conquer strategies. The former reduces theoretical time complexity, but is still memory-heavy. The latter methods sacrifice global context, leading to uninformative and incoherent summaries. This work aims to leverage the memory-efficient nature of divide-and-conquer methods while preserving global context. Concretely, our framework AWESOME uses two novel mechanisms: (1) External memory mechanisms track previously encoded document segments and their corresponding summaries, to enhance global document understanding and summary coherence. (2) Global salient content is further identified beforehand to augment each document segment to support its summarization. Extensive experiments on diverse genres of text, including government reports, transcripts, scientific papers, and novels, show that AWESOME produces summaries with improved informativeness, faithfulness, and coherence than competitive baselines on longer documents, while having a smaller GPU memory footprint.

cs.CL↗

arXiv2023

Dynamics of axion-neutral pseudoscalar mixing

Axions mix with neutral pions after the QCD phase transition through their common coupling to the radiation bath via a Chern-Simons term, as a consequence of the $U(1)$ anomaly. The non-equilibrium effective action that describes this mixing phenomenon is obtained to second order in the coupling of neutral pions and axions to photons. We show that a misaligned axion condensate induces a neutral pion condensate after the QCD phase transition. The dynamics of the pion condensate displays long and short time scales and decays on the longer time scale exhibiting a phenomenon akin to the ``purification'' in a Kaon beam. On the intermediate time scales the macroscopic pion condensate is proportional to a condensate of the abelian Chern-Simons term induced by the axion. We argue that the coupling to the common bath also induces kinetic mixing. We obtain the axion and pion populations, and these exhibit thermalization with the bath. The mutual coupling to the bath induces long-lived axion-neutral pion coherence independent of initial conditions. The framework of the effective action and many of the consequences are more broadly general and applicable to scalar or pseudoscalar particles mixing in a medium.

hep-ph↗

arXiv2023

BUMP: A Benchmark of Unfaithful Minimal Pairs for Meta-Evaluation of Faithfulness Metrics

The proliferation of automatic faithfulness metrics for summarization has produced a need for benchmarks to evaluate them. While existing benchmarks measure the correlation with human judgements of faithfulness on model-generated summaries, they are insufficient for diagnosing whether metrics are: 1) consistent, i.e., indicate lower faithfulness as errors are introduced into a summary, 2) effective on human-written texts, and 3) sensitive to different error types (as summaries can contain multiple errors). To address these needs, we present a benchmark of unfaithful minimal pairs (BUMP), a dataset of 889 human-written, minimally different summary pairs, where a single error is introduced to a summary from the CNN/DailyMail dataset to produce an unfaithful summary. We find BUMP complements existing benchmarks in a number of ways: 1) the summaries in BUMP are harder to discriminate and less probable under SOTA summarization models, 2) unlike non-pair-based datasets, BUMP can be used to measure the consistency of metrics, and reveals that the most discriminative metrics tend not to be the most consistent, and 3) unlike datasets containing generated summaries with multiple errors, BUMP enables the measurement of metrics' performance on individual error types.

cs.CL↗

arXiv2023

Chern Simons condensate from misaligned axions

We obtain the non-equilibrium condensate of the Chern Simons density induced by a misaligned homogeneous coherent axion field in linear response. The Chern-Simons dynamical susceptibility is simply related to the axion self-energy, a result that is valid to leading order in the axion coupling but to all orders in the couplings of the gauge fields to other fields within or beyond the standard model except the axion. The induced Chern-Simons density requires renormalization which is achieved by vacuum subtraction. For ultralight axions of mass $m_a$ coupled to electromagnetic fields with coupling $g$, the renormalized high temperature contribution post-recombination is $\langle \vec{E}\cdot\vec{B}\rangle(t) = -\frac{g\,π^2\,T^4}{15} \,\overline{a}(t)+ \frac{g \,m^2_a\,T}{16\,π}\,\dot{\overline{a}}(t) $ with $\overline{a}(t)$ the dynamical homogeneous axion condensate. We conjecture that emergent axion-like quasiparticle excitations in condensed matter systems may be harnessed to probe cosmological axions and the Chern-Simons condensate. Furthermore, it is argued that a misaligned axion can also induce a non-abelian Chern-Simons condensate of similar qualitative form, and can also ``seed'' chiral symmetry breaking and induce a neutral pion condensate after the QCD phase transition.

hep-ph↗

arXiv2023

Non-equilibrium dynamics of Axion-like particles: the quantum master equation

We study the non-equilibrium dynamics of Axion-like particles (ALP) coupled to Standard Model degrees of freedom in thermal equilibrium. The Quantum Master Equation (QME) for the (ALP) reduced density matrix is derived to leading order in the coupling of the (ALP) to the thermal bath, but to \emph{all} orders of the bath couplings to degrees of freedom within or beyond the Standard Model other than the (ALP). The (QME) describes the damped oscillation dynamics of an initial misaligned (ALP) condensate, thermalization with the bath, decoherence and entropy production within a unifying framework. The (ALP) energy density $\mathcal{E}(t)$ features two components: a ``cold'' component from the misaligned condensate and a ``hot'' component from thermalization with the bath, with $\mathcal{E}(t)= \mathcal{E}_{c}\,e^{-γ(T)\,t}+\mathcal{E}_{h}(1-e^{-γ(T)\,t})$ thus providing a ``mixed dark matter'' scenario. Relaxation of the (ALP) condensate, thermalization, decoherence and entropy production occur on similar time scales. An explicit example with (ALP)-photon coupling, valid post recombination yields a relaxation rate $γ(T)$ with a substantial enhancement from thermal emission and absorption. A misaligned condensate is decaying at least since recombination and on the same time scale thermalizing with the cosmic microwave background (CMB). Possible consequences for birefringence of the (CMB) and (ALP) contribution to the effective number of ultrarelativistic species and galaxy formation are discussed.

astro-ph.CO↗

arXiv2022

Brownian Axion-like particles

We study the non-equilibrium dynamics of a pseudoscalar axion-like particle (ALP) weakly coupled to degrees of freedom in thermal equilibrium by obtaining its reduced density matrix. Its time evolution is determined by the in-in effective action which we obtain to leading order in the (ALP) coupling but to \emph{all orders} in the couplings of the bath to other fields within or beyond the standard model. The effective equation of motion for the (ALP) is a Langevin equation with noise and friction kernels obeying the fluctuation dissipation relation. A ``misaligned'' initial condition yields damped coherent oscillations, however, the (ALP) population increases towards thermalization with the bath. As a result, the energy density features a mixture of a cold component from misalignment and a hot component from thermalization with proportions that vary in time $(cold)\,e^{-Γt}+(hot)\,(1-e^{-Γt})$, providing a scenario wherein the ``warmth'' of the dark matter evolves in time from colder to hotter. As a specific example we consider the (ALP)-photon coupling $g a \vec{E}\cdot \vec{B}$ to lowest order, valid from recombination onwards. For $T \gg m_a$ the long-wavelength relaxation rate is substantially enhanced $Γ_T = \frac{g^2\,m^2_a\,T}{16π} $. The ultraviolet divergences of the (ALP) self-energy require higher order derivative terms in the effective action. We find that at high temperature, the finite temperature effective mass of the (ALP) is $m^2_a(T) = m^2_a(0)\Big[ 1-(T/T_c)^4\Big]$, with $T_c \propto \sqrt{m_a(0)/g}$, \emph{suggesting} the possibility of an inverted phase transition, which when combined with higher derivatives may possibly indicate exotic new phases. We discuss possible cosmological consequences on structure formation, the effective number of relativistic species and birefringence of the cosmic microwave background.

hep-ph↗

arXiv2022

Time-aware Prompting for Text Generation

In this paper, we study the effects of incorporating timestamps, such as document creation dates, into generation systems. Two types of time-aware prompts are investigated: (1) textual prompts that encode document timestamps in natural language sentences; and (2) linear prompts that convert timestamps into continuous vectors. To explore extrapolation to future data points, we further introduce a new data-to-text generation dataset, TempWikiBio, containing more than 4 millions of chronologically ordered revisions of biographical articles from English Wikipedia, each paired with structured personal profiles. Through data-to-text generation on TempWikiBio, text-to-text generation on the content transfer dataset, and summarization on XSum, we show that linear prompts on encoder and textual prompts improve the generation quality on all datasets. Despite having less performance drop when testing on data drawn from a later time, linear prompts focus more on non-temporal information and are less sensitive to the given timestamps, according to human evaluations and sensitivity analyses. Meanwhile, textual prompts establish the association between the given timestamps and the output dates, yielding more factual temporal information in the output.

cs.CL↗

arXiv2022

HIBRIDS: Attention with Hierarchical Biases for Structure-aware Long Document Summarization

Document structure is critical for efficient information consumption. However, it is challenging to encode it efficiently into the modern Transformer architecture. In this work, we present HIBRIDS, which injects Hierarchical Biases foR Incorporating Document Structure into the calculation of attention scores. We further present a new task, hierarchical question-summary generation, for summarizing salient content in the source document into a hierarchy of questions and summaries, where each follow-up question inquires about the content of its parent question-summary pair. We also annotate a new dataset with 6,153 question-summary hierarchies labeled on long government reports. Experiment results show that our model produces better question-summary hierarchies than comparisons on both hierarchy quality and content coverage, a finding also echoed by human judges. Additionally, our model improves the generation of long-form summaries from lengthy government reports and Wikipedia articles, as measured by ROUGE scores.

cs.CL↗

arXiv2021

CLIFF: Contrastive Learning for Improving Faithfulness and Factuality in Abstractive Summarization

We study generating abstractive summaries that are faithful and factually consistent with the given articles. A novel contrastive learning formulation is presented, which leverages both reference summaries, as positive training data, and automatically generated erroneous summaries, as negative training data, to train summarization systems that are better at distinguishing between them. We further design four types of strategies for creating negative samples, to resemble errors made commonly by two state-of-the-art models, BART and PEGASUS, found in our new human annotations of summary errors. Experiments on XSum and CNN/Daily Mail show that our contrastive learning framework is robust across datasets and models. It consistently produces more factual summaries than strong comparisons with post error correction, entailment-based reranking, and unlikelihood training, according to QA-based factuality evaluation. Human judges echo the observation and find that our model summaries correct more errors.

cs.CL↗