811 research outputs found
When Fewer Layers Break More Chains: Layer Pruning Harms Test-Time Scaling in LLMs
No full textLayer pruning has emerged as a widely adopted technique for improving the efficiency of large language models (LLMs). Although existing methods demonstrate strong performance retention on general knowledge tasks, their effect on long-chain reasoning, a more brittle yet crucial capability, remains largely unexplored. In this work, we study the impact of layer pruning on long-chain reasoning through the lens of test-time scaling, a key mechanism in modern LLMs that enables strong reasoning capacity by allocating more computation at inference time. With extensive experiments, we demonstrate that pruning even one or two layers can severely impair test-time scaling, with performance collapsing drastically on long reasoning benchmarks even when performance on knowledge-intensive and shallow reasoning tasks remains stable. Furthermore, we find that standard supervised fine-tuning remedies fail to recover test-time scaling once it has deteriorated. Through in-depth analyses, we identify the mechanisms underlying this fragility of test-time scaling and highlight the fundamental risks of applying layer pruning to reasoning-intensive LLMs. These findings call for a rethinking of layer pruning strategies and provide insights for developing methods that preserve the robustness of reasoning. We open-source the codebase in \href{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}{https://github.com/keyu-wang-2002/Layer-Pruning-Harms-Inference-Scaling}
Synthesis of ultralong MnO/C coaxial nanowires as freestanding anodes for high-performance lithium ion batteries
No full textPublisher's PDFA facile synthesis strategy is reported for the preparation of a freestanding membrane of ultralong MnO/C coaxial nanowires using a novel in situ interfacial polymerization technique. The MnO/C membrane possesses interconnected porous structures with a nanowire diameter of ca. 100 nm and a length of up to hundreds of micrometers. When used as a freestanding anode for lithium ion batteries, the coaxial MnO/C nanocomposites exhibit a high reversible capacity of 832 mA h g−1 at a current density of 100 mA g−1 after 100 cycles, good rate capability and outstanding cycling stability with a specific capacity of 480 mA h g−1 being retained after 600 cycles at a high current density of 1000 mA g−1. The uniform carbon coating formed along the ultralong one-dimensional nanostructure surface is the key-enabling factor that not only improves the electrode reaction kinetics, but also renders excellent cycling performance by accommodating the large volume variation of MnO during charge/discharge processes. The superior electrochemical properties suggest that the facile synthesis strategy can be extended to the fabrication of other freestanding films for potential application in energy storage systems.University of Delaware. Department of Mechanical Engineering
A theoretical study on the statics and dynamics of magnetic domain walls and skyrmions
No full textThis thesis conducts a comprehensive theoretical study and numerical simulations on the statics and dynamics of magnetic domain walls (DWs) and skyrmions, two prominent examples of topological solitons in magnetic materials. For DWs, the focus is on their motion in ferrimagnetic nanowires, driven by external magnetic fields or spin-polarized currents. The research addresses challenges like the Walker breakdown, which limits DW speed. A significant finding is that high-speed DW motion occurs near the angular momentum compensation point (AMCP). The study uses energy conservation principles to explain DW dynamics, proving that static DWs cannot exist under uniform external fields and deriving a velocity formula consistent with experimental results. In the skyrmion section, the thesis investigates ferromagnetic skyrmion pinning by disk-shaped defects and the dynamics of current-driven antiferromagnetic skyrmions in disordered systems. It reveals how skyrmion type and disk size affect pinning and explores the impact of disorder on skyrmion trajectories. Overall, this work enhances the understanding of DWs and skyrmions, offering insights for future spintronic device development based on topological magnetic solitons.</p
Author response image 1. Author response
No full textAnkyrin adaptors together with their spectrin partners coordinate diverse ion channels and cell adhesion molecules within plasma membrane domains and thereby promote physiological activities including fast signaling in the heart and nervous system. Ankyrins specifically bind to numerous membrane targets through their 24 ankyrin repeats (ANK repeats), although the mechanism for the facile and independent evolution of these interactions has not been resolved. Here we report the structures of ANK repeats in complex with an inhibitory segment from the C-terminal regulatory domain and with a sodium channel Nav1.2 peptide, respectively, showing that the extended, extremely conserved inner groove spanning the entire ANK repeat solenoid contains multiple target binding sites capable of accommodating target proteins with very diverse sequences via combinatorial usage of these sites. These structures establish a framework for understanding the evolution of ankyrins' membrane targets, with implications for other proteins containing extended ANK repeat domains.</p
Learn to navigate through deep neural networks
No full textAutonomous navigation is a crucial prerequisite for mobile robots to perform various tasks while it remains a great challenge due to its inherent complexity. This thesis deals with the autonomous navigation problem using deep neural networks. It presents four main parts, i.e., an imitation learning based path planning algorithm, an imitation learning based online path planning method, a deep reinforcement learning based autonomous steering method, and a deep reinforcement learning based autonomous navigation method.
Firstly, as the basis of navigation, path planning has been extensively studied for decades. The computational time of most existing methods depends on environmental conditions, which leads to the compromise between time efficiency and path quality. To address this challenge, a novel end-to-end deep neural network architecture is proposed to learn 3D path planning policies. By embedding the action decomposition and composition concept, the proposed network is capable of generating actions in 3D space merely through 2D convolutional neural networks and exhibits high generalization capability. Moreover, its computational time for each action prediction is almost independent of environmental scale and complexity.
Furthermore, a deep neural network based online path planning method is also proposed. Firstly, an end-to-end network architecture is designed to learn 3D local path planning policies. In addition, a corresponding path planning framework is also developed to achieve real-time online path planning in unknown environments. In the framework, actions are determined efficiently based on the agent's current location, surrounding obstacles and target position. In addition, the efficacy of the planner is further improved through switching among multiple networks considering different environmental ranges. And meanwhile, line-of-sight checks are also performed to optimize the path quality. Without any prior knowledge of the environment, the proposed deep neural network based online planner is competent to generate near-optimal paths in various unknown cluttered environments. Moreover, its computational time and effectiveness are both independent of environmental scale and complexity, which demonstrates its superiority in large-scale complex environments.
Compared to dealing with the path planning problem separately, it is more superior to achieve autonomous navigation via direct mapping of raw sensor data to control commands. In addition, it is also more desirable to learn from past experiences automatically so as to enhance generalization capability in coping with unseen circumstances. Therefore, a deep reinforcement learning algorithm is proposed to achieve autonomous steering in complex environments. The developed model is capable of deriving steering commands from raw depth images in an end-to-end manner. By embedding a branching noisy dueling architecture, the proposed DRL algorithm can learn the autonomous steering policy more effectively while enabling the simultaneous determination of both linear and angular velocities. In addition, a two-stream feature extractor is introduced to improve depth feature extraction through considering temporal variations explicitly. Moreover, a new action selection strategy is also proposed to achieve motion filtering by taking the consistency of angular velocity into account. It is worth noting that the developed model is readily transferable from a simple virtual training environment to various complicated real-world deployments without any fine-tuning.
Furthermore, to address the more challenging goal-directed autonomous navigation problem, a novel deep reinforcement learning algorithm and a tailored network architecture are therefore proposed. The developed model can output control commands directly from raw depth images and encoded destination information, so that the robot can reach the goal positions while avoiding bumping into obstacles. Firstly, a new depth feature extractor is introduced to acquire critical spatiotemporal features from raw depth images. In addition, a double-source scheme is presented to provide more comprehensive learning samples based on a switching criterion. Moreover, a dual network architecture is proposed, which trains two networks associated with different tasks simultaneously. Specifically, the primary network is employed to learn the navigation policy while the auxiliary network is used to learn the depth feature extractor. It is noteworthy that after merely trained in a simple virtual environment, the developed model is readily deployable to a variety of complex real-world scenarios without any fine-tuning.
In summary, this thesis addresses autonomous navigation problems through deep neural networks. Both virtual and real-world experiments have demonstrated the effectiveness and superiority of the proposed methods.Doctor of Philosoph
gen. et sp. nov., a silicified coniferous trunk from the Changhsingian (Permian) of southern Bogda Mountains, northwestern China
No full textA silicified trunk, Zhuotingoxylon liaoi Wan, Yang, Wang, Liu et Wang gen. et sp. nov., is described from the uppermost part of Guodikeng Formation in South Taodonggou section, Turpan-Hami Basin, Xinjiang Uygur Autonomous Region, northwestern China. It is characterized by a solid pith, endarch primary xylem and pycnoxylic wood. The pith is composed of parenchyma and sclereids. Radial walls of primary xylem tracheids have spiral and scalariform thickenings. Secondary xylem consists of thick-walled tracheids and parenchymatous rays. Uniseriate rounded pits with oval apertures are distributed on radial tracheidal walls separately. Cell walls of rays are homogeneous and smooth. Rays are 1-10 cells high in tangential section. Cross-field pits are cupressoid. There are 1-4 bordered pits with slit-like to oval apertures in each cross-field. Based on the anatomical features of the pith and xylems, it is proposed that the new stem has a coniferous affinity. The new fossil stem adds to the knowledge of vascular plant diversity close to the Permian-Triassic boundary
Coaxial MoS2@Carbon Hybrid Fibers: A Low-Cost Anode Material for High-Performance Li-Ion Batteries
No full textPublisher's PDFA low-cost bio-mass-derived carbon substrate has been employed to synthesize MoS2@carbon composites through a hydrothermal method. Carbon fibers derived from natural cotton provide a three-dimensional and open framework for the uniform growth of MoS2 nanosheets, thus hierarchically constructing coaxial architecture. The unique structure could synergistically benefit fast Li-ion and electron transport from the conductive carbon scaffold and porous MoS2 nanostructures. As a result, the MoS2@carbon composites—when serving as anodes for Li-ion batteries—exhibit a high reversible specific capacity of 820 mAh·g−1, high-rate capability (457 mAh·g−1 at 2 A·g−1), and excellent cycling stability. The use of bio-mass-derived carbon makes the MoS2@carbon composites low-cost and promising anode materials for high-performance Li-ion batteries.University of Delaware. Department of Mechanical Engineering
Weyl group twists and representations of quantum affine Borel algebras
No full textWe define categories of representations of Borel subalgebras
of quantum affine algebras
, which come from the category
twisted by Weyl group elements . We construct inductive systems of
finite-dimensional -modules twisted by , which
provide representations in the category . We also establish a
classification of simple modules in these categories .
We explore convergent phenomenon of -characters of representations of
quantum affine algebras, which conjecturally give the -characters of
representations in .
Furthermore, we propose a conjecture concerning the relationship between the
category and the twisted category , and we propose
a possible connection with shifted quantum affine algebras.Comment: 30 page
Systèmes intégrables, algèbres de Borel affines quantiques, et équations aux différences elliptiques
No full textThe subjects of this thesis are organized into three parts. In the first part, we prove a conjecture formulated by E. Frenkel and D. Hernandez. More specifically, we have established the TQ systems and the QQ~ systems for the twisted quantum affine algebras, Langlands dual to the quantum affine algebras. To prove these systems, we develop the representation theory of the Borel subalgebras of twisted quantum affine algebras. In addition, we formulate a conjecture stating a relationship between the Grothendieck ring of the category O of representations of twisted Borel subalgebras and that of the corresponding untwisted type. We establish our conjecture for certain remarkable families of representations. As a result, we deduce the TQ and QQ~ systems for the twisted Borel subalgebras. In the second part, our focus shifts to the untwisted types, where we examine a Weyl group symmetry associated with the category O of representations of Borel subalgebras Uqb. For each element w in W, we study the q-characters normalized by an l-weight labeled by w. We propose a conjecture on the convergence phenomenon of the w-normalized q-characters of finite-dimensional representations. To explain this phenomenon, we define and explore the categories Ow of Uqb-modules, where we classify their simple objects. Specific representations in Ow are constructed as the inductive limit of finite-dimensional representations. We also formulate a conjecture that establishes a link between the usual category O and the new categories Ow. This conjecture suggests a method for calculating the q-characters of a large number of representations in O. In the last part, we redirect our focus to elliptic quantum groups. In this collaborative work with Giovanni Felder and Tommaso Botta, we solve the quantum KZB equations for elliptic quantum groups based on geometric methods. We examine the shuffle structure of the elliptic stable envelopes of Aganagic-Maulik-Okounkov. This shuffle structure allows us to construct solutions to the quantum KZB equations. This work generalizes the construction initially developed by Felder-Tarasov-Varchenko for the specific case of sl2.Les sujets abordés dans cette thèse sont organisés en trois parties. Dans la première partie, nous démontrons une conjecture formulée par E. Frenkel et D. Hernandez. Plus précisément, nous avons prouvé les systèmes TQ et les systèmes QQ~ pour les algèbres affines quantiques tordues, duales de Langlands des algèbres affine quantiques. Pour démontrer ces systèmes, nous développons la théorie des représentations des sous-algèbres de Borel des algèbres affines quantiques tordues. Nous proposons de plus une conjecture énonçant une relation entre l'anneau de Grothendieck de la catégorie O des représentations des sous-algèbres de Borel tordues et celui du type non tordu correspondant. Nous établissons notre conjecture pour une certaines familles de représentations. Nous en déduisons des systèmes TQ et QQ~ pour les sous-algèbres de Borel tordues. Dans la deuxième partie de la thèse, notre attention se porte sur les types non tordus, où nous examinons une symétrie du groupe de Weyl W associée à la catégorie O des représentations des sous-algèbres de Borel Uqb. Pour chaque élément w dans W, nous étudions les q-caractères normalisés par un l-poids marqué par w. Nous proposons une conjecture sur le phénomène de convergence des q-caractères w-normalisés des représentations de dimension finie. Pour expliquer ce phénomène, nous définissons et explorons les catégories Ow de Uqb-modules, où nous classifions leurs objets simples. Des représentations spécifiques dans Ow sont construites comme la limite inductive de représentations de dimension finie. Nous formulons également une conjecture qui établit un lien entre la catégorie usuelle O et les nouvelles catégories Ow. Cette conjecture suggère une méthode pour calculer les q-caractères d'un grand nombre de représentations dans O. Dans la dernière partie de la thèse, nous orientons notre attention vers les groupes quantiques elliptiques. Dans ce travail collaboratif avec Giovanni Felder et Tommaso Botta, nous résolvons les équations KZB quantiques pour les groupes quantiques elliptiques avec des méthodes géométriques. Nous étudions la structure de mélange des enveloppes stables elliptiques d'Aganagic-Maulik-Okounkov. Cette structure de mélange nous permet de construire des solutions aux équations KZB quantiques. Ce travail généralise la construction initialement élaborée par Felder-Tarasov-Varchenko pour le cas spécifique de sl2
- …
