arXiv subjects

Zhihong Lu

Publications and source records attributed to Zhihong Lu.

15 recordsLinked to original sources

arXiv2026

What Matters for Aggressive Decoding-Time KV Eviction? Temporal Aggregation and Ranking Preservation

Decoding-time KV cache compression research focuses heavily on designing better token scoring functions, while the temporal rule that aggregates scores across decode steps is often treated as an implementation detail. Under aggressive KV compression, we find that exponential-moving-average (EMA) aggregation makes approximately order-preserving scorer modifications largely indistinguishable at the eviction-set level. Value-norm and entropy variants remain highly correlated with attention and produce nearly unchanged retention sets, whereas KeyDiff, key norm, recency, and a learned scorer alter the ranking and degrade substantially. We associate this stability with the evaluated aggregation, which couples layer weighting and temporal retention. Building on this observation, we introduce InertiaKV, an EMA-based decoding-time eviction method, and InertiaKV-Lazy, its periodic-refresh variant, which yields 1.34-1.46x decode throughput relative to full refresh InertiaKV. We also study Score-Free decoding as a separate empirical operating point: it scores the full context once at the first decode step, freezes that ranking, and incurs an average quality change of +0.03 while removing all subsequent scoring. Across six open-weight backbones and the LongBench, LongBench-v2, and RULER benchmarks, the results identify temporal aggregation and ranking preservation as distinct, consequential design factors; they do not imply that scoring quality is irrelevant in general.

cs.AI↗

arXiv2026

Anisotropic Spin Polarization and magnetic spin hall effect in Ferromagnets

Spin-dependent transport in ferromagnets underpins the development of high-density spintronic memories. Spin-dependent transport in strong spin-orbit-coupled ferromagnets exhibits a significant anisotropy. Both the overall spin polarization during charge transport and the magnetic spin Hall conductivity are found to exhibit pronounced anisotropy when the magnetization is tilted away from the crystallographic easy axis or when the electric field is rotated relative to the crystal axes. These anisotropic responses originate primarily from spin-orbit coupling, which is identified as the key driver of the large anisotropy observed in ferromagnet. Furthermore, strain tunability of the magnetic spin Hall anisotropy is demonstrated, with tensile strain progressively enhancing the oscillatory amplitude of the spin Hall conductivity. These findings establish strong spin-orbit-coupled ferromagnets as a platform for anisotropic spin-current generation and field-free spintronic devices that exploit intrinsic material anisotropy for improved performance and energy efficiency.

cond-mat.mes-hall↗

arXiv2026

SimpleSearch-VL: A Simple Recipe for Multimodal Agentic Deep Search

We present SimpleSearch-VL, an efficient, reliable, and practical framework for multimodal agentic search. Its core idea is to improve the agent's own search-and-verification process rather than scaling data, tools, or auxiliary model components. For efficiency, Factorized Adaptive Rollout (FAR) improves sampling efficiency by forming more informative training groups while using redundant samples to mitigate long-tail latency and expose hard samples. For reliability, SimpleSearch-VL performs evidence-verified reasoning, explicitly using chain-of-thought verification to assess the relevance of retrieved visual and textual cues to the original context. For practicality, SimpleSearch-VL keeps a lightweight tool interface and performs webpage self-summary within the agent, requiring no additional external dependencies. With only 5K supervised tool-interleaved trajectories and 2K RL data, SimpleSearch-VL improves Qwen3-VL agentic baselines by 15.8 and 16.0 average points for the 8B and 30B-A3B variants, respectively. The SimpleSearch-VL-30B-A3B model further achieves performance competitive with agentic Gemini-3-Pro.

cs.CV↗

arXiv2026

Nearly Complete Charge--Spin Conversion via Strain-Eliminated Fermi Pockets in $d$-Wave Altermagnets

Ideal $d$-wave altermagnets with nearly orthogonal flat Fermi surfaces enable complete spin-channel separation and 100% theoretical charge--spin conversion efficiency (CSE). The metallic altermagnet $\mathrm{KV_2Se_2O}$ exemplifies this, but realistic samples host residual elliptical Fermi pockets that enhance charge conductivity while suppressing spin conductivity, drastically reducing CSE. Here we show that in-plane equibiaxial tensile strain systematically eliminates these parasitic pockets, restoring the flat-band geometry. Our first-principles calculations reveal that CSE increases monotonically with strain, reaching a record $\sim$96% at 4% strain. An effective tight-binding model confirms that pocket suppression, governed by reduced next-nearest-neighbor hoppings, is the dominant mechanism. We further identify an unconventional out-of-plane spin current component with CSE $\sim$55% at optimal orientations, enabling field-free perpendicular magnetization switching. Moreover, the same strain-driven removal of parasitic pockets yields giant TMR enhancement in $\mathrm{KV_2Se_2O}$-based magnetic tunnel junctions, from $10^{5}%$ to $10^{9}%$, and the giant TMR persists over a wide energy window near the Fermi level. These findings establish strain engineering as a clean, widely applicable strategy to maximize CSE and magnetoresistance in $d$-wave altermagnets, and provide predictive descriptors for screening high-efficiency spintronic materials.

cond-mat.mtrl-sci↗

arXiv2026

DisSR: Disentangling Speech Representation for Degradation-Prior Guided Cross-Domain Speech Restoration

Previous speech restoration (SR) primarily focuses on single-task speech restoration (SSR), which cannot address general speech restoration problems. Training specific SSR models for different distortions is time-consuming and lacks generality. In addition, most studies ignore the problem of model generalization across unseen domains. To overcome those limitations, we propose DisSR, a Disentangling Speech Representation based general speech restoration model with two properties: 1) Degradation-prior guidance, which extracts speaker-invariant degradation representation to guide the diffusion-based speech restoration model. 2) Domain adaptation, where we design cross-domain alignment training to enhance the model's adaptability and generalization on cross-domain data, respectively. Experimental results demonstrate that our method can produce high-quality restored speech under various distortion conditions. Audio samples can be found at https://itspsp.github.io/DisSR.

cs.SD↗

arXiv2026

PulseMind: A Multi-Modal Medical Model for Real-World Clinical Diagnosis

Recent advances in medical multi-modal models focus on specialized image analysis like dermatology, pathology, or radiology. However, they do not fully capture the complexity of real-world clinical diagnostics, which involve heterogeneous inputs and require ongoing contextual understanding during patient-physician interactions. To bridge this gap, we introduce PulseMind, a new family of multi-modal diagnostic models that integrates a systematically curated dataset, a comprehensive evaluation benchmark, and a tailored training framework. Specifically, we first construct a diagnostic dataset, MediScope, which comprises 98,000 real-world multi-turn consultations and 601,500 medical images, spanning over 10 major clinical departments and more than 200 sub-specialties. Then, to better reflect the requirements of real-world clinical diagnosis, we develop the PulseMind Benchmark, a multi-turn diagnostic consultation benchmark with a four-dimensional evaluation protocol comprising proactiveness, accuracy, usefulness, and language quality. Finally, we design a training framework tailored for multi-modal clinical diagnostics, centered around a core component named Comparison-based Reinforcement Policy Optimization (CRPO). Compared to absolute score rewards, CRPO uses relative preference signals from multi-dimensional com-parisons to provide stable and human-aligned training guidance. Extensive experiments demonstrate that PulseMind achieves competitive performance on both the diagnostic consultation benchmark and public medical benchmarks.

cs.CV↗

arXiv2025

High-Efficiency Nonrelativistic Charge-Spin Conversion in X-Type Antiferromagnets

Antiferromagnetic materials with spin splitting have attracted considerable attention for their symmetry-enabled anisotropic spin textures that sustain a zero net magnetization, thereby facilitating efficient spin-current generation. In this work, the highly efficient generation of nonrelativistic spin currents is demonstrated to be facilitated by the distinctive Fermi surface geometry of X-type collinear antiferromagnets. As a prototype conducting X-type antiferromagnet, the Fermi surface of $\beta-\mathrm{Fe_2PO_5}$ exhibits a distinct $d$-wave altermagnetic characteristic, which compresses into a nearly X-shaped configuration. This results in highly efficient spin currents, achieving a charge-spin conversion efficiency of up to 90\%. Moreover, the spin current polarization is controlled by the orientation of the N\'eel vector. When the N\'eel vector tilts to the out-of-plane direction, an in-plane injected charge current can generate a special spin current component with both spin polarization and propagation along the out-of-plane direction, whose charge-spin conversion efficiency substantially exceeds that of known ferromagnets, altermagnets, noncollinear antiferromagnets, and low-symmetry materials. The highly efficient charge-spin conversion in X-type antiferromagnets provides a novel and highly effective spin source system for the development of low-power spintronic devices.

cond-mat.mtrl-sci↗

arXiv2025

Strain-Induced Non-alter Compensated Magnet and Its Application to Magnetic Tunnel Junction Device Design

The recent proposal of altermagnetism has drawn widespread attention to antiferromagnet (AFM) exhibiting spin splitting, extending beyond the realm of sign-alternating spin splitting in momentum space protected solely by rotational symmetry. Herrin, we propose a shear-strain strategy that enables significant modulation of d-wave altermagnets into an non-alter compensated magnets. A comprehensive analysis combining the magnetic moment compensation characteristics of opposite spin sublattices with the distribution of spin-resolved conduction channels in momentum space under the [001] crystal orientation reveals that shear strain breaks the rotational symmetry of alternatmagnets. To explore the application potential of non-alter compensated magnets, we designe RuO2/TiO2/RuO2 magnetic tunnel junctions (MTJ) with three crystallographic orientations ((001), (110), (100)) and investigated their transport properties under shear strain. This non-alter electronic structure not only enhances the tunneling magnetoresistance (TMR) in spin-split paths of intrinsic RuO2 (226% to 431%) but also enables substantial TMR in spin-degenerate paths (from 0-88%)). Our work provides guidelines for broadening magnetic materials and device platforms.

cond-mat.mtrl-sci↗

arXiv2025

Strain-induced nonrelativistic altermagnetic spin splitting effect

Recent studies reveal that $\mathcal{T}$-odd spin currents generated via the nonrelativistic altermagnetic spin splitting effect (ASSE) exhibit significant potential for spintronics applications, with both computational and experimental validations. Addressing the scarcity of conductive altermagnets, we propose strain engineering as a reliable method for inducing altermagnetism. Focusing on rutile-structured $\mathrm{OsO}_2$, first-principles calculations show that minor equibiaxial tensile strain ($\mathcal{E}_{\mathrm{ts}}$=3\%) induces nonmagnetic-to-altermagnetic transitions, achieving an ASSE-driven spin-charge conversion ratio ($\theta_{\text{AS}}$) of $\sim$7\% -- far surpassing conventional spin Hall angles ($\theta_{\text{IS}}$). Calculations reveal that substantial $\theta_{\text{AS}}$ persists even in the absence of spin-orbit coupling, with its magnitude positively correlating to nonrelativistic spin splitting magnitude, which further confirms the strain-induced ASSE's nonrelativistic origin. Further investigation reveals that $\mathrm{RuO}_2$ exhibits analogous phenomena, which may resolve recent controversies regarding its magnetic properties. Our research opens new simple pathways for developing next-generation altermagnetic spintronic devices.

cond-mat.mtrl-sci↗

arXiv2025

AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society

Understanding human behavior and society is a central focus in social sciences, with the rise of generative social science marking a significant paradigmatic shift. By leveraging bottom-up simulations, it replaces costly and logistically challenging traditional experiments with scalable, replicable, and systematic computational approaches for studying complex social dynamics. Recent advances in large language models (LLMs) have further transformed this research paradigm, enabling the creation of human-like generative social agents and realistic simulacra of society. In this paper, we propose AgentSociety, a large-scale social simulator that integrates LLM-driven agents, a realistic societal environment, and a powerful large-scale simulation engine. Based on the proposed simulator, we generate social lives for over 10k agents, simulating their 5 million interactions both among agents and between agents and their environment. Furthermore, we explore the potential of AgentSociety as a testbed for computational social experiments, focusing on five key social issues: polarization, the spread of inflammatory messages, the effects of universal basic income policies, the impact of external shocks such as hurricanes, and urban sustainability. These five issues serve as valuable cases for assessing AgentSociety's support for typical research methods -- such as surveys, interviews, and interventions -- as well as for investigating the patterns, causes, and underlying mechanisms of social issues. The alignment between AgentSociety's outcomes and real-world experimental results not only demonstrates its ability to capture human behaviors and their underlying mechanisms, but also underscores its potential as an important platform for social scientists and policymakers.

cs.SI↗

arXiv2025

Emergence of polarization in networks of large language model agents

Rapid advances in large language models (LLMs) have not only empowered autonomous agents to generate social networks, communicate, and form shared and diverging opinions on political issues, but have also begun to play a growing role in shaping human political deliberation. Our understanding of their collective behaviours and underlying mechanisms remains incomplete, however, posing unexpected risks to human society. In this paper, we simulate networked systems involving thousands of LLM agents across different backbone models (GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3), in which agents interact through LLM-guided conversations and update their opinions over time, resulting in the emergence of opinion polarization. We discover that these agents spontaneously develop their own social network with properties characteristic of human social networks, including homophilic clustering. The collective opinions of these LLM agents evolve in ways that exhibit behavioural patterns consistent with social phenomena and mechanisms widely discussed in studies of human behaviour. Overall, these behaviours, patterns, and emergent phenomena produced by LLM agents are consistent with real-world observations and established opinion-dynamics models. This consistency suggests that LLM agents can serve as a valuable synthetic testbed for exploring hypothetical intervention strategies in networked LLM-agent systems.

cs.SI↗

arXiv2024

Extremely Large Anisotropy of Effective Gilbert Damping in Half-Metallic CrO2

Half-metals are a class of quantum materials with 100% spin-polarization at the Fermi level and have attracted a lot of attention for future spintronic device applications. CrO2 is one of the most promising half-metal candidates, for which the electrical and magnetic properties have been intensively studied in the last several decades. Here, we report the observation of a giant anisotropy (~1600%) of effective Gilbert damping in the single crystalline half metallic (100)-CrO2 thin films, which is significantly larger than the values observed on conventional ferromagnetic Fe and CoFe thin films. Furthermore, the effective Gilbert damping exhibits opposite temperature-dependent behaviors below 50 K with magnetic field along [010] direction and near [001] direction. These experimental results suggest the strong spin-orbit coupling anisotropy of the half-metallic CrO2 and might pave the way for future magnonic computing applications.

cond-mat.mtrl-sci↗

arXiv2023

S$^3$: Social-network Simulation System with Large Language Model-Empowered Agents

Social network simulation plays a crucial role in addressing various challenges within social science. It offers extensive applications such as state prediction, phenomena explanation, and policy-making support, among others. In this work, we harness the formidable human-like capabilities exhibited by large language models (LLMs) in sensing, reasoning, and behaving, and utilize these qualities to construct the S$^3$ system (short for $\textbf{S}$ocial network $\textbf{S}$imulation $\textbf{S}$ystem). Adhering to the widely employed agent-based simulation paradigm, we employ prompt engineering and prompt tuning techniques to ensure that the agent's behavior closely emulates that of a genuine human within the social network. Specifically, we simulate three pivotal aspects: emotion, attitude, and interaction behaviors. By endowing the agent in the system with the ability to perceive the informational environment and emulate human actions, we observe the emergence of population-level phenomena, including the propagation of information, attitudes, and emotions. We conduct an evaluation encompassing two levels of simulation, employing real-world social network data. Encouragingly, the results demonstrate promising accuracy. This work represents an initial step in the realm of social network simulation empowered by LLM-based agents. We anticipate that our endeavors will serve as a source of inspiration for the development of simulation systems within, but not limited to, social science.

cs.SI↗

arXiv2019

Voltage-controlled skyrmion-based artificial synapse in a synthetic antiferromagnet

Spintronics exhibits significant potential in neuromorphic computing system with high speed, high integration density, and low dissipation. In this letter, we propose an ultralow-dissipation spintronic memristor composed of a synthetic antiferromagnet (SAF) and a piezoelectric substrate. Skyrmions/skyrmion bubbles can be generated in the upper layer of SAF with weak anisotropy energy (Ea). With a weak electric field on the heterostructure, the interlayer antiferromagnetic coupling can be manipulated, giving rise to a continuous transition between a large skyrmion bubble and a small skyrmion. This thus induces the variation of the resistance of a magnetic tunneling junction. The synapse based on this principle may manipulate the weight in a wide range at a cost of a very low energy consumption of 0.3 fJ. These results pave a way to ultralow power neuromorphic computing applications.

physics.comp-ph↗

arXiv2007

Domain wall switching: optimizing the energy landscape

It has recently been suggested that exchange spring media offer a way to increase media density without causing thermal instability (superparamagnetism), by using a hard and a soft layer coupled by exchange. Victora has suggested a figure of merit xi = 2 E_b/mu_0 m_s H_sw, the ratio of the energy barrier to that of a Stoner-Wohlfarth system with the same switching field, which is 1 for a Stoner-Wohlfarth (coherently switching) particle and 2 for an optimal two-layer composite medium. A number of theoretical approaches have been used for this problem (e.g., various numbers of coupled Stoner-Wohlfarth layers and continuum micromagnetics). In this paper we show that many of these approaches can be regarded as special cases or approximations to a variational formulation of the problem, in which the energy is minimized for fixed magnetization. The results can be easily visualized in terms of a plot of the energy as a function of magnetic moment m_z, in which both the switching field [the maximum slope of E(m_z)] and the stability (determined by the energy barrier E_b) are geometrically visible. In this formulation we can prove a rigorous limit on the figure of merit xi, which can be no higher than 4. We also show that a quadratic anistropy suggested by Suess et al comes very close to this limit.

cond-mat.stat-mech↗