251 research outputs found
Analytical investigations of nonlinear stiffness characteristics of Halbach-cylinder magnetic springs for heavy-load capacity
No full textQuasi-zero stiffness (QZS) has become a promising way of realizing low-frequency vibration isolation, where magnetic springs have been widely adopted for constructing negative stiffness. However, existing single-layer magnetic springs often have a small-amplitude negative stiffness, so the loading capacity is low. In order to address this issue, this paper presents novel Halbach-cylinder magnetic springs (HCMSs) by using the Halbach array. Firstly, stiffness formulas of basic single-layer magnetic springs are analytically built based on the Amperian current model. The stiffness of the HCMS is derived from combining multiple single-layer magnetic springs. Then, nonlinear stiffness characteristics of both single-layer magnetic springs and HCMSs are investigated in terms of the amplitude, the uniformity, and the displacement range of negative stiffness. Analytical results show that HCMSs can generate negative stiffness with different equilibrium positions, and the amplitude of negative stiffness of HCMSs is much larger than that of single-layer magnetic springs. The amplitude of negative stiffness is in conflict with the uniformity, so a trade-off design is needed. In addition, increasing the number of layers of Halbach cylinders can be adopted to realize larger-amplitude and wider-range negative stiffness. This study will provide new insights into designing QZS with heavy-load capacity
Scalable foundation models
No full textThe continual growth in computational resources and human annotators, driven by advances in hardware architecture and the increasing accessibility of crowdsourcing platforms, has created unprecedented opportunities for artificial intelligence (AI) models. However, without scalable AI solutions (e.g., training algorithms, model architectures), much of this additional compute and annotation may be underutilized or yield diminishing returns. Furthermore, as real-world applications demand increasingly sophisticated AI capabilities, scalable models offer a clear path to achieving higher levels of intelligence by taking full advantage of available computational and annotation resources. This makes scalability not just a technical consideration, but a fundamental requirement for advancing the field of AI in parallel with hardware developments.
This dissertation investigates the fundamental trajectory toward scalable foundation models through three subsequent research milestones. (1) Predictable Scaling: We examine scaling laws that govern the development of foundation models, analyzing how models' capabilities correlate with computational resources. Our research establishes principle solutions to forecasting model behaviors and resource requirements across different scales, enabling scientific and reliable scaling of AI models. (2) Scalable Modeling: We explore model architectures and training recipes optimized for multimodal learning, demonstrating how these approaches can effectively utilize increasing computational resources and data to achieve continuous performance improvements. Our findings reveal architectural principles and training strategies that maintain efficiency at scale while avoiding common bottlenecks in previous modeling strategies. (3) Scalable Oversight: We study the scalable post-training approaches that enable continuous model improvement and alignment with human values even as model capabilities expand beyond human expertise. This research introduces novel techniques for scalable supervision that scale in parallel with model complexity and capability, ensuring the responsible advancement of AI models. In Chapter 1, we describe the key research problems and dive deeply into several key featured research in the following chapters.
Predictable Scaling
In Chapter 2, we study how to estimate the actual capabilities (i.e., downstream performance) in large language models (LLMs) via addressing the challenges of LLMs' emergent abilities. We focus on the pre-training loss as a more computation-efficient metric for performance estimation. We present FLP, a two-stage approach for performance prediction that consists of first estimating a function that maps computational resources (e.g., FLOPs) to the pre-training Loss using a series of sampling models, followed by mapping the pre-training loss to downstream task Performance after the critical "emergent phase".
Scalable Modeling
In Chapter 3, we present a scalable code-guided visual representation learning method and a single transformer architecture for scalable vision-language modeling. A single unified Transformer architecture can effectively addresses the scalability concerns in previous large vision-language models (LVLMs); however, its limited adoption in modern context likely stems from the absence of reliable training recipes that balance both modalities and ensure stable training for billion-scale models. We introduce the first open-source training recipe for developing unified LVLMs, using moderate academic resources (8 x A100 80GB GPUs). In addition, we revisit the next token prediction loss on vision-language pre-training, and argue that this can be a false proxy of the actual capabilities in LVLMs. We propose a new algorithm, ViStruct, to scale up vision-langugage pre-training. The results show that ViStruct scales better with more data and compute.
Scalable Oversight
In Chapter 4, we investigate a novel approach to AI supervision through learning from AI feedback. We introduce a scalable alignment framework that harnesses the strong capabilities of large language models (LLMs) to guide the development of LVLMs. Our framework advances beyond conventional numerical reward signals by leveraging natural language feedback as a primary mechanism for model optimization and refinement. This methodology enables the systematic refinement of model responses, promoting attributes of helpfulness, truthfulness, and safety and also enhance their capacity for sustained multi-turn interactions. Our approach demonstrates how advanced LLMs can serve as effective supervisors in the training pipeline, offering a scalable solution to the challenge of model alignment. In addition, we utilize the AI feedback to supervise the reasoning consistency of LVLMs. In our curated benchmark that targets the chain-of-thought (CoT) reasoning performance and consistency of LVLMs, the results show that supervising the reasoning process brings better reasoning capabilities in LVLMs.Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2026-02-19 without embargo termsThe student, Yangyi Chen, accepted the attached license on 2025-12-02 at 14:33.The student, Yangyi Chen, submitted this Dissertation for approval on 2025-12-02 at 14:46.This Dissertation was approved for publication on 2025-12-03 at 10:14.DSpace SAF Submission Ingestion Package generated from Vireo submission #23022 on 2026-02-19 at 18:26:1
Advancing Multi-Scale Geographic Environmental Monitoring: A Synthesis of Cutting-Edge Research and Scalable Solutions
No full textThe geographic environment is a complex concept that encompasses various natural elements of the Earth’s surface and human activities [...
A new <i>Rattus</i> species and its associated micromammals from the Pliocene Yangyi Formation in Baoshan, western Yunnan, China
No full textRattus (sensu stricto) is one of the groups of rodents and is most closely related to human activities. The diversity of extant Rattus species is the highest among rodents, but fossil species are rare, mainly found in Asia during the Late Pliocene to Early Pleistocene. Here, we described a Late Pliocene new species of Rattus based on material from the upper Yangyi Formation in Baoshan county, Yunnan Province, China. The primitive morphological characteristics of fewer molar roots, t12 preserved on M1–2, t9 preserved on M3, and buccal stylids well developed on lower molars indicate that the new Rattus is by far the most primitive form in the Rattus genus. In dental morphology of Rattus, the molars have an evolutionary trend of gradually increasing the number of roots and fusing cusps into an arched ridge, reduction to disappearance of the t7 and t12 on upper molars, and reduction to disappearance of the mesiocentral cusp on m1, the anterior buccal stylid on m1–3 and hypoconid on m3. Discovery of the new Rattus from Baoshan confirms that Rattus probably originated in southern Asia before the Late Pliocene, and also provides new fossil evidence to calibrate the molecular clock of the divergence of Rattus from other murines. In addition, we also reported five associated small mammals: Neotetracus sp. nov., Anourosorex qianensis, Soricidae gen. et sp. indet., Ia io, and Kowalskia sp. Most of them inhabit warm and humid tropical or subtropical montane forests or shrubland environments. The composition of this mammalian assemblage indicates that the paleoelevation and paleoclimate of Baoshan area in the Pliocene are very similar to those of the present.</p
Monsoon versus uplift in southwestern China : Late Pliocene climate in Yuanmou Basin, Yunnan
No full textYuanmou Basin of Yunnan, SW China, is a famous locality with hominids, hominoids, mammals and plant fossils. Based on the published megaflora and palynoflora data from Yuanmou Basin, the climate of Late Pliocene is reconstructed using the Coexistence Approach. The results indicate a warm and humid subtropical climate with a mean annual temperature of ca. 16–17°C and a mean annual precipitation of ca. 1500–1600 mm in the Late Pliocene rather than a dry, hot climate today, which may be due to the local tectonic change and gradual intensification of India monsoon. The comparison of Late Pliocene climate in Eryuan, Yangyi, Longling, and Yuanmou Basin of Yunnan Province suggests that the mean annual temperatures generally show a latitudinal gradient and fit well with their geographic position, while the mean annual precipitations seem to be related to the different geometries of the valleys under the same monsoon system
Wang Meng and contemporary Chinese literature: the vicissitudes of a committed writer
Get PDFThis thesis examines the way Wang Meng has developed as a writer from the 1950s to the 1990s in the context of New China's political and literary background. It looks at the compromises he was forced to make between his political beliefs in the Communist Party and his chosen role as a professional writer. After his disastrous early foray into what was deemed to be unacceptable political criticism with The Young Newcomer in the Organisation Department in the 1950s, when the opportunity came to start publishing again in the late 1970s he was boldly innovative in style, helping to transform New Period literature, but conservative in content, sticking to politically acceptable topics. It was only with Hard Porridge in 1989 that he ventured again, and very successfully, into political comment. There is no outstanding leading writer in contemporary China, but Wang Meng is a leading contender for the title
On the impact of the COVID-19 pandemic on the household’s consumption and labor supply: theory and application
Get PDFThe COVID-19 pandemic and the corresponding regulation measures carried out to curb it have had a strong negative impact on the whole economy, and household consumption has been seriously affected. A large part of the drop in consumption is due to the reduction of household income, which is mainly caused by the labor supply loss during the pandemic. To present the mechanism of the impact of the pandemic on consumption, this study constructs a novel theoretical model. Two hypotheses about the pandemic’s impact on labor supply are proposed and empirically tested. Subsequently, a comparative static analysis is carried out to determine the numerical mechanism of the pandemic’s impact on household consumption. In addition, the model is also empirically tested and further modified for application, enabling the studies of both a realistic simulation and a policy simulation. This study finds that the labor supply of households has been affected during the pandemic, and there is a mediating effect channel through the regulation stringency. The epidemic severity and regulation policies have a negative impact on household consumption, in turn, will raise the saving rate of households. The income effect of the two on consumption accounts for 32% and 44% of the total effect respectively.
First published online 05 September 202
Parallel Jacobian-free Newton Krylov discrete ordinates method for pin-by-pin neutron transport models
Get PDFA parallel Jacobian-Free Newton Krylov discrete ordinates method (comePSn_JFNK) is proposed to solve the multi-dimensional multi-group pin-by-pin neutron transport models, which makes full use of the good efficiency and parallel performance of the JFNK framework and the high accuracy of the Sn method for the large-scale models. In this paper, the k-eigenvalue and the scalar fluxes (rather than the angular fluxes) are chosen as the global solution variables of the parallel JFNK method, and the corresponding residual functions are evaluated by the Koch–Baker–Alcouffe (KBA) algorithm with the spatial domain decomposition in the parallel Sn framework. Unlike the original Sn iterative strategy, only a “flattened” power iterative process which includes a single outer iteration without nested inner iterations is required for the JFNK strategy. Finally, the comePSn_JFNK code is developed in C++ language and, the numerical solutions of the 2-D/3-D KAIST-3A benchmark problems and the 2-D/3-D full-core MOX/UOX pin-by-pin models with different control rod distribution show that comePSn_JFNK method can obtain significant efficiency advantage compared with the original power iteration method (comePSn) for the parallel simulation of the large-scale complicated pin-by-pin models
ANALISIS TATA RUANG KOLEKSI YANG OPTIMAL BAGI KENYAMANAN PEMUSTAKA DI PERPUSTAKAAN UNIVERSITAS PGRI PALEMBANG
No full textLibraries are built to support the achievement of goals by providing collections or library materials to be utilized by students, lecturers, and library staff. Activities in this library are based on the PKL (Field Work Practice) program of the Library Science Study Program, Faculty of Adab and Humanities, Raden Fatah State Islamic University Palembang. This activity is intended so that library staff know how to organize the optimal space for the comfort and satisfaction of the library users, it is also expected that from this activity the library users can more easily find the information or library materials needed. In the implementation of the activity, the author used the interview method and direct observation to the library by directly observing the state of the library layout, this activity was assisted by other PKL students who were also grouped for PKL, and also assisted by one of the officers on duty there. Based on the results of direct research, on the 3rd floor is a circulation room, librarian workspace, reading room, and collection room. Then in the arrangement of the room on the 3rd floor is quite optimal where this library has paid attention to various factors of good and correct library room arrangemen
Jacobian-Free Newton Krylov Methods for Steady and Transient Neutron Transport Models
No full textThere is an urgent need to reduce the computational costs and improve the convergence rate for the three-dimensional (3D) high-fidelity pin-by-pin full core simulation. Therefore, the efficient and robust acceleration method of 3D large-scale pin-by-pin neutron transport models is a primary objective of high-resolution reactor calculations. In this paper, comeSn_JFNK, an efficient unified parallel solver was developed for 3D steady-state and transient pin-by-pin neutron transport models. The comeSn_JFNK solver integrated the parallel discrete ordinate (SN) neutron transport code comeSn into the parallel computational framework comeJFNK of Jacobian-Free Newton Krylov (JFNK). The comeSn code and comeJFNK framework were developed by the Virtual Reactor Coupling Analysis Team (VRCAT) at Huazhong University of Science and Technology (HUST). Therefore, comeSn_JFNK took advantages of the fast and robust convergence of the parallel JFNK framework and the high accuracy and efficiency of the SN method based on the KBA algorithm. To further improve computational efficiency of comeSn_JFNK, in the JFNK solution, neutron scalar fluxes instead of neutron angular fluxes were chosen as the global solution variables. This can reduce the number of variables and minimize the computing scale. To rapidly construct the unified JFNK residuals, the parallel KBA transport sweep methods and physics-based preconditioning techniques were utilized to improve the computational efficiency. To rapidly construct the unified JFNK residuals, the parallel KBA transport sweep methods and physics-based preconditioning techniques were utilized. This way of constructing residuals, for both steady-state and transient models, can further improve the computational efficiency. Finally, the detailed analysis of the computational accuracy and acceleration characteristics of comeSn_JFNK was presented by solving the pin-by-pin steady-state KAIST-3A and homogenized-pin transient C5G7-TD2 benchmark cases. The relative errors of radial average power density, effective multiplication factor and relative powers as a function of time were shown in this paper. There are almost the same numerical solutions between the parallel comeSn code and the parallel solver called comSn_JFNK. Numerical results also show that the parallel solver called comeSn_JFNK can achieve an acceleration of over 10 times for KAIST-3A benchmark problems and approximately 30 times for C5G7-TD2 cases compared to the original parallel comeSn code using the traditional source iteration/power iteration methods. These solutions indicate that the JFNK methods offer significant acceleration for solving both steady-state and transient SN neutron transport models compared to traditional source iteration/power iteration methods. In summary, the JFNK method provides reliable computational accuracy and high computational efficiency, which demonstrates the potential and advantages to accelerate neutron transport solutions. It also establishes a foundation for efficient simultaneous solution of more complicated pin-by-pin neutron transport and thermal-hydraulic coupling problems in the nuclear reactor cores using the unified JFNK framework in the coupling multiphysics environment (COME) developed by VRCAT at HUST
- …
