arXiv subjects

Zhenyan Lu

Publications and source records attributed to Zhenyan Lu.

9 recordsLinked to original sources

arXiv2026

Fictional Worldbuilding: Multi-Agent LLM Collaboration with Hierarchical Context Compression and Iterative Review

Worldbuilding, the construction of coherent fictional worlds, is a foundational task in game design and literary creation. Large Language Models (LLMs) offer new possibilities for automated content generation, but their application to worldbuilding faces three challenges: context explosion that grows linearly with the building process, the tension between creative diversity and content consistency, and the absence of automated quality assurance. This paper presents AutoWorldBuilder, a multi-agent collaborative system that addresses these challenges through five integrated components: a structured concept network with conflict detection; a DAG-based hybrid batch scheduler that groups tasks by semantic locality; a four-layer context compression mechanism achieving approximately 90% token reduction; an iterative review system with specialized Auditor agents that improves proposal pass rates from 42% to over 85%; and a skill-driven agent architecture supporting zero-code extension with differentiated temperature configuration. Two experiments across 20 diverse worldbuilding tasks, using GPT-OSS 120B and DeepSeek v3.2 as LLM backends, demonstrate a 95.0% success rate. The system generated 56-103 self-consistent concepts per world in 18-31 minutes with zero-conflict delivery. The architectural patterns validated here, including layer-as-budget compression, semantic-locality scheduling, and separation of generation and review, transfer to the broader class of knowledge-intensive, multi-agent LLM applications.

cs.AI↗

arXiv2025

MobiEdit: Resource-efficient Knowledge Editing for Personalized On-device LLMs

Large language models (LLMs) are deployed on mobile devices to power killer applications such as intelligent assistants. LLMs pre-trained on general corpora often hallucinate when handling personalized or unseen queries, leading to incorrect or outdated responses. Knowledge editing addresses this by identifying and adjusting a small crucial portion of model weights, without compromising the general knowledge. However, prior knowledge editing methods are impractical to run on local devices due to the resource-heavy backpropagation (BP) needed for updates. We present MobiEdit, the first mobile knowledge editing framework that enables efficient LLM personalization on commercial off-the-shelf (COTS) mobile devices. MobiEdit replaces full-precision BP with quantized forward-only gradient estimation, thus compatible with the energy-efficient mobile neural processing units (NPUs). MobiEdit replaces full-precision backpropagation with quantized forward-only gradient estimation, making it compatible with energy-efficient mobile NPUs. To further improve gradient estimation efficiency, we introduce two optimizations: an early stoping mechanism that adaptively terminates editing upon success and a prefix cache that reuses computation across steps. Our approach enables real-time editing of a 3B-parameter model (Qwen2.5-3B-Instruct) on COTS mobile devices with 7.6$\times$ less memory, 14.7 $\times$ less energy and 3.6$\times$ less latency compared to previous knowledge editing methods.

cs.LG↗

arXiv2025

The influence of cosmological constant on light deflection in rotating spacetimes via the generalized Gibbons-Werner method

Recently, we proposed a generalized Gibbons-Werner (GW) method for analyzing particle trajectories in rotating spacetimes, regardless of their asymptotic behavior [Huang \textit{et al.}, \href{https://iopscience.iop.org/article/10.1088/1475-7516/2024/01/013}{J. Cosmol. Astropart. Phys. 01(2024), 013}]. Using this method, we examine the impact of the cosmological constant ($\Lambda$) on the light deflection in rotating spacetimes within the framework of Kerr-de Sitter (KdS) geometry. Although Sultana previously calculated the deflection angle of light in KdS spacetime, our study advances this research in three aspects: (i) Orbit solution -- the light trajectory is derived by directly solving the original equation of motion (EOM) without intermediate processes. (ii) Positions of the source and observer -- the finite distances of the source and observer from the lens are explicitly considered, avoiding approximations. (iii) Staticity of the source and observer -- the Randers optical space is employed to resolve the staticity constraint. Through these refined considerations, we obtain a novel expression for the deflection angle of light in KdS spacetime, accurate to second-order in $\Lambda$, as well as in the mass (M) and spin parameter (a) of the central body. Furthermore, we discuss the discrepancies between our results and previous expressions. Finally, we evaluate the observational implications of our corrections relative to Sultana's work in the lensing systems of the Sun and Sgr A*, and show that they may become observable with forthcoming high-precision astronomical measurements.

gr-qc↗

arXiv2024

Ultrafast demagnetization in ferromagnetic materials: Origins and progress

Since the discovery of ultrafast demagnetization in Ni thin films in 1996, laser-induced ultrafast spin dynamics have become a prominent research topic in the field of magnetism and spintronics. This development offers new possibilities for the advancement of spintronics and magnetic storage technology. The subject has drawn a substantial number of researchers, leading to a series of research endeavors. Various models have been proposed to elucidate the physical processes underlying laser-induced ultrafast spin dynamics in ferromagnetic materials. However, the potential origins of these processes across different material systems and the true contributions of these different origins remain challenging in the realm of ultrafast spin dynamics. This predicament also hinders the development of spintronic terahertz emitters. In this review, we initially introduce the different experimental methods used in laser-induced ultrafast spin dynamics. We then systematically explore the magnetization precession process and present seven models of ultrafast demagnetization in ferromagnetic materials. Subsequently, we discuss the physical processes and research status of four ultrafast demagnetization origins (including spin-flipping, spin transport, non-thermal electronic distribution, and laser-induced lattice strain). Since attosecond laser technique and antiferromagnetic materials exhibit promising applications in ultrahigh-frequency spintronics, we acknowledge the emerging studies used by attosecond pules and studies on ultrafast spin dynamics in antiferromagnets, noting the significant challenges that need to be addressed in these burgeoning field.

cond-mat.mtrl-sci↗

arXiv2024

PhoneLM:an Efficient and Capable Small Language Model Family through Principled Pre-training

The interest in developing small language models (SLM) for on-device deployment is fast growing. However, the existing SLM design hardly considers the device hardware characteristics. Instead, this work presents a simple yet effective principle for SLM design: architecture searching for (near-)optimal runtime efficiency before pre-training. Guided by this principle, we develop PhoneLM SLM family (currently with 0.5B and 1.5B versions), that acheive the state-of-the-art capability-efficiency tradeoff among those with similar parameter size. We fully open-source the code, weights, and training datasets of PhoneLM for reproducibility and transparency, including both base and instructed versions. We also release a finetuned version of PhoneLM capable of accurate Android Intent invocation, and an end-to-end Android demo. All materials are available at https://github.com/UbiquitousLearning/PhoneLM.

cs.CL↗

arXiv2024

Small Language Models: Survey, Measurements, and Insights

Small language models (SLMs), despite their widespread adoption in modern smart devices, have received significantly less academic attention compared to their large language model (LLM) counterparts, which are predominantly deployed in data centers and cloud environments. While researchers continue to improve the capabilities of LLMs in the pursuit of artificial general intelligence, SLM research aims to make machine intelligence more accessible, affordable, and efficient for everyday tasks. Focusing on transformer-based, decoder-only language models with 100M-5B parameters, we survey 70 state-of-the-art open-source SLMs, analyzing their technical innovations across three axes: architectures, training datasets, and training algorithms. In addition, we evaluate their capabilities in various domains, including commonsense reasoning, mathematics, in-context learning, and long context. To gain further insight into their on-device runtime costs, we benchmark their inference latency and memory footprints. Through in-depth analysis of our benchmarking data, we offer valuable insights to advance research in this field.

cs.CL↗

arXiv2024

A Survey of Resource-efficient LLM and Multimodal Foundation Models

Large foundation models, including large language models (LLMs), vision transformers (ViTs), diffusion, and LLM-based multimodal models, are revolutionizing the entire machine learning lifecycle, from training to deployment. However, the substantial advancements in versatility and performance these models offer come at a significant cost in terms of hardware resources. To support the growth of these large models in a scalable and environmentally sustainable way, there has been a considerable focus on developing resource-efficient strategies. This survey delves into the critical importance of such research, examining both algorithmic and systemic aspects. It offers a comprehensive analysis and valuable insights gleaned from existing literature, encompassing a broad array of topics from cutting-edge model architectures and training/serving algorithms to practical system designs and implementations. The goal of this survey is to provide an overarching understanding of how current approaches are tackling the resource challenges posed by large foundation models and to potentially inspire future breakthroughs in this field.

cs.LG↗

arXiv2024

Testing the cosmic distance duality relation with Type Ia supernova and transverse BAO measurements

In this work, we test the cosmic distance duality relation (CDDR) by comparing the angular diameter distance (ADD) derived from the transverse Baryon Acoustic Oscillations (BAO) data with the luminosity distance (LD) from the Pantheon type Ia supernova (SNIa) sample. The binning method and Gaussian process are employed to match ADD data with LD data at the same redshift. First, we use nonparametric and parametric methods to investigate the impact of the specific prior values of the absolute magnitude $M_{\rm B}$ from SNIa observations and the sound horizon scale $r_{\rm s}$ from transverse BAO measurements on the CDDR tests. The results obtained from the parametric and non-parametric methods indicate that specific prior values of $M_{\rm B}$ and $r_{\rm s}$ lead to significant biases on the CDDR test. Then, to avoid these biases, we propose a method independent of $M_{\rm B}$ and $r_{\rm s}$ to test CDDR by considering the fiducial value of $κ\equiv10^{M_{\rm B} \over 5}r_{\rm s}$ as a nuisance parameter and then marginalizing its influence with a flat prior in the analysis. No violation of the CDDR is found, and the transverse BAO measurement can be used as a powerful tool to verify the validity of CDDR in the cosmological-model-independent method.

astro-ph.CO↗

arXiv2024

Generalized Gibbons-Werner method for stationary spacetimes

The Gibbons-Werner (GW) method is a powerful approach in studying the gravitational deflection of particles moving in curved spacetimes. The application of the Gauss-Bonnet theorem (GBT) to integral regions constructed in a two-dimensional manifold enables the deflection angle to be expressed and calculated from the perspective of geometry. However, different techniques are required for different scenarios in the practical implementation which leads to different GW methods. For the GW method for stationary axially symmetric (SAS) spacetimes, we identify two problems: (a) the integral region is generally infinite, which is ill-defined for some asymptotically nonflat spacetimes whose metric possesses singular behavior, and (b) the intricate double and single integrals bring about complicated calculation, especially for highly accurate results and complex spacetimes. To address these issues, a generalized GW method is proposed in which the infinite region is replaced by a flexible region to avoid the singularity, and a simplified formula involving only a single integral of a simple integrand is derived by discovering a significant relationship between the integrals in conventional methods. Our method provides a comprehensive framework for describing the GW method for various scenarios. Additionally, the generalized GW method and simplified calculation formula are applied to three different kinds of spacetimes--Kerr spacetime, Kerr-like black hole in bumblebee gravity, and rotating solution in conformal Weyl gravity. The first two cases have been previously computed by other researchers, affirming the effectiveness and superiority of our approach. Remarkably, the third case is newly examined, yielding an innovative result for the first time.

gr-qc↗