arXiv subjects

Zhiyuan Li

Publications and source records attributed to Zhiyuan Li.

At least 163 records · Page 9Linked to original sources

arXiv2025

Exploiting Local Observations for Robust Robot Learning

While many robotic tasks can be addressed using either centralized single-agent control with full state observation or decentralized multi-agent control, clear criteria for choosing between these approaches remain underexplored. This paper systematically investigates how multi-agent reinforcement learning (MARL) with local observations can improve robustness in complex robotic systems compared to traditional centralized control. Through theoretical analysis and empirical validation, we show that in certain tasks, decentralized MARL can achieve performance comparable to centralized methods while exhibiting greater resilience to perturbations and agent failures. By analytically demonstrating the equivalence of single-agent reinforcement learning (SARL) and MARL under full observability, we identify observability as the critical factor distinguishing the two paradigms. We further derive bounds quantifying performance degradation under external perturbations for locally observable policies. Empirical results on standard MARL benchmarks confirm that MARL with limited observations can maintain competitive performance. Finally, real-world experiments with a mobile manipulator demonstrate that decentralized MARL controllers achieve markedly improved robustness to agent malfunctions and environmental disturbances relative to centralized baselines. Together, these findings highlight MARL with local observations as a robust and practical alternative to conventional centralized control in complex robotic systems.

cs.RO↗

arXiv2025

On the convergence of PINNs for inverse source problem in the complex Ginzburg-Landau equation

This paper addresses the problem of recovering the spatial profile of the source in the complex Ginzburg-Landau equation from regional observation data at fixed times. We establish two types of sufficient measurements for the unique solvability of the inverse problem. The first is to determine the source term by using whole data at one fixed instant. Conditional stability is established by using the eigenfunction expansion argument. Next, using the analytic continuation method, both uniqueness and a stability estimate for recovering the unknown source can be established from local data at two instants. Finally, algorithms based on the physics-informed neural networks (PINNs) are proposed, and several numerical experiments are presented to show the accuracy and efficiency of the algorithm.

math.AP↗

arXiv2025

Megrez2 Technical Report

We present Megrez2, a novel lightweight and high-performance language model architecture optimized for device native deployment. Megrez2 introduces a novel cross-layer expert sharing mechanism, which significantly reduces total parameter count by reusing expert modules across adjacent transformer layers while maintaining most of the model's capacity. It also incorporates pre-gated routing, enabling memory-efficient expert loading and faster inference. As the first instantiation of the Megrez2 architecture, we introduce the Megrez2-Preview model, which is pre-trained on a 5-trillion-token corpus and further enhanced through supervised fine-tuning and reinforcement learning with verifiable rewards. With only 3B activated and 7.5B stored parameters, Megrez2-Preview demonstrates competitive or superior performance compared to larger models on a wide range of tasks, including language understanding, instruction following, mathematical reasoning, and code generation. These results highlight the effectiveness of the Megrez2 architecture to achieve a balance between accuracy, efficiency, and deployability, making it a strong candidate for real-world, resource-constrained applications.

cs.CL↗

arXiv2025

Precise Measurement of $^{216}$Po Half-life with Exact Parent-daughter Pairing in PandaX-4T

We report a precise measurement of $^{216}\rm Po$ half-life using the PandaX-4T liquid xenon time projection chamber (TPC). $^{220}\rm Rn $, emanating from a $^{228}\rm Th $ calibration source, is injected to the detector and undergoes successive $α$ decays, first to $^{216}\rm Po$ and then to $^{212}\rm Pb$. PandaX-4T detector measures the 5-dimensional (5D) information of each decay, including time, energy, and 3-dimensional positions. Therefore, we can identify the $^{220}\rm Rn $ and $^{216}\rm Po$ decay events and pair them exactly to extract the lifetime of each $^{216}\rm Po$. With a large data set and high-precision $^{220}\rm $Rn-$^{216}\rm $Po pairing technique, we measure the $^{216}\rm Po$ half-life to be $143.7\pm0.5$ ms, which is the most precise result to date and agrees with previously published values. The leading precision of this measurement demonstrates the power of 5D calorimeter and the potential of exact parent-daughter pairing in the xenon TPC.

nucl-ex↗

arXiv2025

Structured Preconditioners in Adaptive Optimization: A Unified Analysis

We present a novel unified analysis for a broad class of adaptive optimization algorithms with structured (e.g., layerwise, diagonal, and kronecker-factored) preconditioners for both online regret minimization and offline convex optimization. Our analysis not only provides matching rate to several important structured preconditioned algorithms including diagonal AdaGrad, full-matrix AdaGrad, and AdaGrad-Norm, but also gives an improved convergence rate for a one-sided variant of Shampoo over that of original Shampoo. Interestingly, more structured preconditioners (e.g., diagonal Adagrad, AdaGrad-Norm which use less space and compute) are often presented as computationally efficient approximations to full-matrix Adagrad, aiming for improved optimization performance through better approximations. Our unified analysis challenges this prevailing view and reveals, perhaps surprisingly, that more structured preconditioners, despite using less space and computation per step, can outperform their less structured counterparts. To demonstrate this, we show that one-sided Shampoo, which is relatively much cheaper than full-matrix AdaGrad could outperform it both theoretically and experimentally.

cs.LG↗

arXiv2025

Wind-fed Supermassive Black Hole Accretion by the Nuclear Star Cluster: the Case of M31*

The central supermassive black hole (SMBH) of the Andromeda galaxy, known as M31*, exhibits dim electromagnetic emission and is inferred to have an extremely low accretion rate for its remarkable mass ($\sim10^8~\rm~M_\odot$). In this work, we use three-dimensional hydrodynamical simulations to explore a previously untested scenario, in which M31* is fed by the collective stellar mass-loss from its surrounding nuclear star cluster, manifested as a famous eccentric disk of predominantly old stellar populations. The stellar mass-loss is assumed to be dominated by the slow and cold winds from 100 asymptotic giant-branch stars, which follow well-constrained Keplerian orbits around M31* and together provide a mass injection rate of $\sim4\times10^{-5}\rm~M_\odot~yr^{-1}$. The simulations achieve a quasi-steady state on a Myr timescale, at which point a quasi-Keplerian, cool ($T\sim10^3-10^4~\rm K$) gas disk extending several parsecs is established. This disk is continuously supplied by the stellar winds and itself feeds the central SMBH. At the end of the simulations at 2 Myr, an accretion rate of $\sim2\times10^{-5}\rm~M_\odot~yr^{-1}$ is found but could vary by a factor of few depending on whether the subdominant gravity of the NSC or a moderate global inflow is included. The predicted X-ray luminosity of $\sim10^{36}~\rm erg~s^{-1}$, dominated by the hot ($T\sim10^7-10^8~\rm K$) plasma within 0.2 parsec of the SMBH, is well consistent with Chandra observations. We conclude that the feeding mechanism of M31* is successfully identified, which has important implications for the working of dormant SMBHs prevalent in the local universe.

astro-ph.GA↗

arXiv2025

External Large Foundation Model: How to Efficiently Serve Trillions of Parameters for Online Ads Recommendation

Ads recommendation is a prominent service of online advertising systems and has been actively studied. Recent studies indicate that scaling-up and advanced design of the recommendation model can bring significant performance improvement. However, with a larger model scale, such prior studies have a significantly increasing gap from industry as they often neglect two fundamental challenges in industrial-scale applications. First, training and inference budgets are restricted for the model to be served, exceeding which may incur latency and impair user experience. Second, large-volume data arrive in a streaming mode with data distributions dynamically shifting, as new users/ads join and existing users/ads leave the system. We propose the External Large Foundation Model (ExFM) framework to address the overlooked challenges. Specifically, we develop external distillation and a data augmentation system (DAS) to control the computational cost of training/inference while maintaining high performance. We design the teacher in a way like a foundation model (FM) that can serve multiple students as vertical models (VMs) to amortize its building cost. We propose Auxiliary Head and Student Adapter to mitigate the data distribution gap between FM and VMs caused by the streaming data issue. Comprehensive experiments on internal industrial-scale applications and public datasets demonstrate significant performance gain by ExFM.

cs.IR↗

arXiv2025

Multi-Modality Conditioned Variational U-Net for Field-of-View Extension in Brain Diffusion MRI

An incomplete field-of-view (FOV) in diffusion magnetic resonance imaging (dMRI) can severely hinder the volumetric and bundle analyses of whole-brain white matter connectivity. Although existing works have investigated imputing the missing regions using deep generative models, it remains unclear how to specifically utilize additional information from paired multi-modality data and whether this can enhance the imputation quality and be useful for downstream tractography. To fill this gap, we propose a novel framework for imputing dMRI scans in the incomplete part of the FOV by integrating the learned diffusion features in the acquired part of the FOV to the complete brain anatomical structure. We hypothesize that by this design the proposed framework can enhance the imputation performance of the dMRI scans and therefore be useful for repairing whole-brain tractography in corrupted dMRI scans with incomplete FOV. We tested our framework on two cohorts from different sites with a total of 96 subjects and compared it with a baseline imputation method that treats the information from T1w and dMRI scans equally. The proposed framework achieved significant improvements in imputation performance, as demonstrated by angular correlation coefficient (p < 1E-5), and in downstream tractography accuracy, as demonstrated by Dice score (p < 0.01). Results suggest that the proposed framework improved imputation performance in dMRI scans by specifically utilizing additional information from paired multi-modality data, compared with the baseline method. The imputation achieved by the proposed framework enhances whole brain tractography, and therefore reduces the uncertainty when analyzing bundles associated with neurodegenerative.

cs.CV↗

arXiv2025

Robust Containerization of the High Angular Resolution Functional Imaging (HARFI) Pipeline

Historically, functional magnetic resonance imaging (fMRI) of the brain has focused primarily on gray matter, particularly the cortical gray matter and associated nuclei. However, recent work has demonstrated that functional activity in white matter also plays a meaningful role in both cognition and learning. In previous work, we introduced the High Angular Resolution Functional Imaging (HARFI) pipeline, which demonstrated both local and global patterns of functional correlation in white matter. Notably, HARFI enabled exploration of asymmetric voxel-wise correlation using odd-order spherical harmonics. Although the original implementation of HARFI was released via GitHub, adoption was limited due to the technical complexity of running the source code. In this work, we present a robust and efficient containerized version of the HARFI pipeline, enabling seamless execution across multiple public datasets. Our goal is to facilitate broader and deeper exploration of functional white matter architecture, especially through the lens of high angular resolution functional correlations. The key innovation of this work is the containerized implementation, which we have made available under a permissive open-source license to support reproducible and accessible research practices.

physics.med-ph↗

arXiv2025

N-body simulations of galaxy bars generated by satellite collisions: effects of the impact geometry

Bars, a common and important structure in disk galaxies, can be induced by galaxy interactions. Although there have been some studies on bar formation in flybys or collisions, the vast parameter space still leaves many scenarios that require further investigation. Here, we focus on the role of collisions caused by small galaxies (denoted as intruders), referred to as satellite collisions, in bar formation for MW/M31-like galaxies (denoted as target galaxies). Multiple sets of simulations with varying intruder initial velocities, inclination angles, collision positions, and intruder masses were run to study the dependence of this mechanism on these parameters. Our simulations show that bar formation favors moderate collision velocity, large inclination angle, off-center collision position, and large intruder mass. However, the bar's pattern speed and length are insensitive to these parameters, as the intruder's mass is relatively small compared to that of the target galaxy itself. Moreover, based on our tests, the intruder mass should be more than $\sim 3\times10^{9}$ ${\rm M}_{\odot}$ in order for this bar formation mechanism to operate effectively in MW/M31-like galaxies. Our work suggests the possibility that satellite collisions may have contributed, to some extent, to the bar formation in the Milky Way or M31.

astro-ph.GA↗

arXiv2025

PENCIL: Long Thoughts with Short Memory

While state-of-the-art LLMs have demonstrated great promise of using long Chains-of-Thought (CoT) to boost reasoning, scaling it up to more challenging problems at test-time is fundamentally limited by suboptimal memory usage -- intermediate computations accumulate indefinitely in context even when no longer needed for future thoughts. We introduce PENCIL, which incorporates a novel reduction mechanism into the autoregressive generation process that recursively cleans up intermediate thoughts based on patterns learned from training. By iteratively generating and erasing thoughts, PENCIL can think deeper to solve harder problems using shorter context and less compute. Empirically, we observe PENCIL is significantly more effective and efficient than CoT. For example, we demonstrate PENCIL with a small 25M-parameter transformer and 2048 context length solves Einstein's puzzle -- a task that challenges much larger models like GPT-4. Theoretically, we prove PENCIL can perform universal efficient computation by simulating any Turing machines with optimal time and space complexity, and thus can solve arbitrary computable tasks that are otherwise intractable for vanilla CoT.

cs.LG↗

arXiv2025

Find a Scapegoat: Poisoning Membership Inference Attack and Defense to Federated Learning

Federated learning (FL) allows multiple clients to collaboratively train a global machine learning model with coordination from a central server, without needing to share their raw data. This approach is particularly appealing in the era of privacy regulations like the GDPR, leading many prominent companies to adopt it. However, FL's distributed nature makes it susceptible to poisoning attacks, where malicious clients, controlled by an attacker, send harmful data to compromise the model. Most existing poisoning attacks in FL aim to degrade the model's integrity, such as reducing its accuracy, with limited attention to privacy concerns from these attacks. In this study, we introduce FedPoisonMIA, a novel poisoning membership inference attack targeting FL. FedPoisonMIA involves malicious clients crafting local model updates to infer membership information. Additionally, we propose a robust defense mechanism to mitigate the impact of FedPoisonMIA attacks. Extensive experiments across various datasets demonstrate the attack's effectiveness, while our defense approach reduces its impact to a degree.

cs.CR↗

arXiv2025

Derived isogenies and isogenies for abelian surfaces

In this paper, we study the twisted Fourier-Mukai partners of abelian surfaces. Following the work of Huybrechts [doi:10.4171/CMH/465], we introduce the twisted derived equivalence between abelian surfaces. We show that there is a twisted derived Torelli theorem for abelian surfaces over algebraically closed fields with characteristic $\neq 2,3$. Over complex numbers, the derived isogenies correspond to rational Hodge isometries between the second cohomology groups, which is in analogy to the work of Huybrechts and Fu-Vial on K3 surfaces. Their proof relies on the global Torelli theorem over $\mathbb{C}$, which is missing in positive characteristics. To overcome this issue, we firstly extend a trick given by Shioda on integral Hodge structures, to rational Hodge structures, $\ell$-adic Tate modules and $F$-crystals. Then we make use of Tate's isogeny theorem to give a characterization of the twisted derived equivalences between abelian surfaces via isogenies. As a consequence, we show the two abelian surfaces are principally isogenous if and only if they are derived isogenous.

math.AG↗

arXiv2025

Numerical analysis of scattered point measurement-based regularization for backward problems for fractional wave equations

In this work, our aim is to reconstruct the unknown initial value from terminal data. We develop a numerical framework on nonuniform time grids for fractional wave equations under the lower regularity assumptions. Then, we introduce a regularization method that effectively handles scattered point measurements contaminated with stochastic noise. The optimal error estimates of stochastic convergence not only balance discretization errors, the noise, and the number of observation points, but also propose an a priori choice of regularization parameters. Finally, several numerical experiments are presented to demonstrate the efficiency and accuracy of the algorithm.

math.NA↗

arXiv2025

Scattered point measurement-based regularization for backward problems for fractional wave equations

In this work, we are devoted to the reconstruction of an unknown initial value from the terminal data. The asymptotic and root-distribution properties of Mittag-Leffler functions are used to establish stability of the backward problem. Furthermore, we introduce a regularization method that effectively handles scattered point measurements contaminated with stochastic noise. Furthermore, we prove the stochastic convergence of our proposed regularization and provide an iterative algorithm to find the optimal regularization parameter. Finally, several numerical experiments are presented to demonstrate the efficiency and accuracy of the algorithm.

math.NA↗

arXiv2025

Filtrations on the derived category of twisted K3 surfaces

We introduce and study the Shen-Yin-Zhao filtration on derived categories of twisted K3 surfaces. A main contribution is the construction of a twisted Beauville-Voisin class $\mathfrak{o}_{\mathscr{X}} \in \operatorname{CH}_0(X)$ that extends fundamental results of O'Grady and Shen-Yin-Zhao \cite{OG13, SYZ20} to twisted settings. This class enables: 1. A derived equivalence-invariant filtration $\mathbf{S}_\bullet(\mathrm{D}^{(1)}(\mathscr{X}))$ preserved under Fourier-Mukai transforms, 2. A birational invariant filtration $\mathbf{S}^{\mathrm{SYZ}}_\bullet \operatorname{CH}_0$ on Bridgeland moduli spaces. We prove $\mathbf{S}^{\mathrm{SYZ}}_\bullet \operatorname{CH}_0$ coincides with Voisin's filtration $\mathbf{S}^{\mathrm{BV}}_\bullet\operatorname{CH}_0$ (Theorem 1.4), providing a canonical candidate for the conjectural Beauville-Voisin filtration. Applications include Bloch's conjecture for (anti)-symplectic automorphisms and existence of algebraically coisotropic subvarieties.

math.AG↗

arXiv2025

The Picard group of the Baily--Borel compactification of the moduli space of quasi-polarized K3 surfaces and generalizations

In this paper, we investigate the Picard group of the Baily--Borel compactification of orthogonal Shimura varieties. As a key result, we determine the Picard group of the Baily--Borel compactification of the moduli space of quasi-polarized K3 surfaces, proving that it is isomorphic to $\mathbb{Z}$. Notably, this contrasts with the moduli space of smooth curves, where the Picard group exhibits a more complex structure after natural compactification. Our result follows from a general theorem for orthogonal Shimura varieties: for even lattices $M$ of signature $(2,n)$ with $n > 8$ satisfying specific arithmetic conditions (e.g. K3 type or $p$-elementary), the rational Picard group of $\overline{\operatorname{Sh}}_Γ(M)$ with $Γ$ containing the stable orthogonal group is $1$-dimensional. The core of our proof lies in constructing an arithmetic obstruction space that governs the extension of Heegner divisors to the generic points of the boundary in orthogonal Shimura varieties. We further establish a connection between this obstruction space and the space of theta series, demonstrating that the obstruction space is maximal under our conditions.

math.AG↗

arXiv2025

Adam Exploits $\ell_\infty$-geometry of Loss Landscape via Coordinate-wise Adaptivity

Adam outperforms SGD when training language models. Yet this advantage is not well-understood theoretically -- previous convergence analysis for Adam and SGD mainly focuses on the number of steps $T$ and is already minimax-optimal in non-convex cases, which are both $\widetilde{O}(T^{-1/4})$. In this work, we argue that the exploitation of nice $\ell_\infty$-geometry is the key advantage of Adam over SGD. More specifically, we give a new convergence analysis for Adam under novel assumptions that loss is smooth under $\ell_\infty$-geometry rather than the more common $\ell_2$-geometry, which yields a much better empirical smoothness constant for GPT-2 and ResNet models. Our experiments confirm that Adam performs much worse when the favorable $\ell_\infty$-geometry is changed while SGD provably remains unaffected. We also extend the convergence analysis to blockwise Adam under novel blockwise smoothness assumptions.

cs.LG↗