arXiv subjects

Yu Tian

Publications and source records attributed to Yu Tian.

At least 73 records · Page 4Linked to original sources

arXiv2026

Internal structure of Hayward black holes

Regular black holes, free of central singularities, provide an ideal laboratory for probing the geometric structure of spacetime. The global structure of some regular black holes, e.g. Hayward black hole, features an event horizon and a Cauchy horizon, raising fundamental questions about the latter's stability. In this work, we investigate collapse of a scalar field in Hayward spacetime. Under weak scalar perturbations, the inner horizon maintains a stable finite radius. In the circumstance of a strong scalar field, the inner horizon shrinks to zero volume, accompanied by the formation of a spacelike singularity. The Hayward geometry is effectively converted into a Schwarzschild-like geometry. Furthermore, the strength of the scalar field governs the contraction dynamics of the inner horizon. As the parameter $p$ of the initial profile for the scalar field approaches the critical threshold ${p_*}$, the radius of the inner horizon ${r_{-}}$ exhibits a universal scaling behavior: ${r_{-}}\propto{|p - {p_*}|^γ}$, with a critical exponent $γ\approx 0.5$.

gr-qc↗

arXiv2026

Tidal Love numbers and the dynamical instability of AdS bubbles

In this work, we study non-radial perturbations of AdS bubbles and their tidal Love numbers (TLNs). The odd- and even-parity TLNs are computed up to $l=6$ in the limit $k \to \infty$. The odd-parity TLNs are found to be negative, while the even-parity TLNs are positive for $\upsilon^2_s=-1$. As $l$ increases, the tidal Love numbers approach zero. The TLNs of the even-parity sector up to order $l=41$ are also calculated over the entire parameter space of $k$, from $0$ to $\infty$. We find that in the region where $p/σ>0$, an increasing number of TLNs become negative as $l$ increases. For $l = 41$, the highest order we have examined, the TLNs are negative everywhere except in a narrow region very close to the zero of $p/σ$, which agrees well with the instability criterion in the eikonal limit for self-gravitating membranes proposed by Yang {\it et al.}\ [P. R. L. {\bf 130}, 011402 (2023)].

gr-qc↗

arXiv2026

Generalizing Perron--Frobenius theory and eigenvector-based centralities to networks with complex edge weights

A fundamental concept in linear algebra and its applications to network analysis is the Perron--Frobenius (PF) theorem, which underpins eigenvector-based centrality measures such as eigenvector centrality, PageRank, and hubs and authorities. By invoking the PF theorem, we know for strongly connected networks with positive edge weights that the eigenvector corresponding to the largest eigenvalue of the weight matrix yields a well-defined centrality measure (namely, eigenvector centrality). Traditional formulations of the PF theorem and associated centrality measures assume that networks have real-valued weights. However, many networks in areas such as quantum information, quantum chemistry, electrodynamics, and machine learning have complex-valued edge weights. In this paper, we study generalizations of the PF theorem to complex-valued matrices, establish connections between these generalizations, and propose generalized eigenvector-based centrality measures to analyzing node importances in networks with complex edge weights. We also prove results about the existence of complex-weighted networks that satisfy generalized PF properties and calculate associated centrality measures for several examples, which we draw from application areas such as electron transport, circuit analysis, mathematical chemistry, and communication networks.

math.SP↗

arXiv2026

GEMINI: Generalized Ensnarlment Measure from Incomplete-linkage of Network-network Interactions

Spatially embedded networks are central to many physical and biological systems, where geometry and connectivity jointly shape structure and function. Examples abound across the scales of biological organization, from network-like membrane-bound organelles in the cell to mesoscale tissue organization of multiple distinct flow networks in organs and beyond. In each of these cases, the complexity of the architectures has heretofore frustrated our ability to link mechanism or regulation of these structures to reduced modeling or even relevant characterization, putting structure-function relationships largely out of reach. Complex, functional spatial networks can be decomposed into tree-like and cyclic substructures, but we still lack both an understanding of how these elements intertwine to give rise to function, and the tools to holistically quantify both the topological and geometric aspects of these features in their full network context. To close this gap, we here introduce GEMINI, a topology and geometry aware operator that directly characterizes incomplete linking and more general spatial associations between edges in spatially embedded network architectures. GEMINI contains information on edge-edge association through an incomplete version of the Gauss linking integral which simultaneously endows it with topological sensitivity when collections of edges form linked assemblages. Validation on both synthetic lattices and on mouse brain vasculature data demonstrates that GEMINI systematically captures and classifies the complexity of structural organizations. Our results provide a general approach for analyzing spatial networks in realistic data, where topology and geometry together determine function, thus opening the door to a more complete understanding of structure-function relationships across a broad set of biological examples where complex network organization is key.

physics.comp-ph↗

arXiv2026

Safe-FedLLM: Delving into the Safety of Federated Large Language Models

Federated learning (FL) addresses privacy and data-silo issues in the training of large language models (LLMs). Most prior work focuses on improving the efficiency of federated learning for LLMs (FedLLM). However, security in open federated environments, particularly defenses against malicious clients, remains underexplored. To investigate the security of FedLLM, we conduct a preliminary study to analyze potential attack surfaces and defensive characteristics from the perspective of LoRA updates. We find two key properties of FedLLM: 1) LLMs are vulnerable to attacks from malicious clients in FL, and 2) LoRA updates exhibit distinct behavioral patterns that can be effectively distinguished by lightweight classifiers. Based on these properties, we propose Safe-FedLLM, a probe-based defense framework for FedLLM, which constructs defenses across three levels: Step-Level, Client-Level, and Shadow-Level. The core concept of Safe-FedLLM is to perform probe-based discrimination on each client's local LoRA updates, treating them as high-dimensional behavioral features and using a lightweight classifier to determine whether they are malicious. Extensive experiments demonstrate that Safe-FedLLM effectively improves FedLLM's robustness against malicious clients while maintaining competitive performance on benign data. Notably, our method effectively suppresses the impact of malicious data without significantly affecting training speed, and remains effective even under high malicious client ratios.

cs.CR↗

arXiv2026

OncoTriad-QA: A Patient-Level Radiology-Pathology-Genomics Benchmark for Pan-Cancer Reasoning

Cancer diagnosis and characterization require integrating complementary evidence from radiology, pathology, genomics, and clinical metadata. However, most medical large language model (LLM) and vision-language model (VLM) benchmarks focus on isolated modalities or narrow image-text tasks, leaving patient-level oncology assessment across multiple evidence streams largely untested. We introduce OncoTriad-QA, a patient-level radiology-pathology-genomics benchmark for pan-cancer question answering. OncoTriad-QA contains 86.1k semantic questions across 9,281 TCGA patient cases from 32 cancer cohorts, aligning CT/MRI radiology, whole-slide histopathology, somatic mutations, copy-number alterations, DNA methylation, bulk RNA-seq, and clinical metadata. Case-specific annotations are constructed through a source-grounded LLM-assisted pipeline using curated labels, diagnostic reports, molecular profiles, and modality-derived evidence as primary sources of truth, with automated consistency checks and clinician review. We also introduce OncoVLM, a reference multimodal model that maps modality-native radiology, pathology, DNA methylation, and RNA-seq evidence into an LLM interface through learned projectors. Experiments show that existing general-purpose and medical LLMs remain limited on comprehensive pan-cancer QA, especially when questions require integrating imaging findings, tumor morphology, and molecular evidence. After fine-tuning on OncoTriad-QA, OncoVLM exceeds MedGemma-4B by an average of 10.7 points when using MCQ accuracy and BERTScore-F1, with consistent gains across multiple-choice and open-ended questions under radiology-only, pathology-only, and all-available settings. These results demonstrate the benchmark's value for training and evaluating models for integrated cancer question answering.

cs.CL↗

arXiv2026

Can Oscillatory and Persistent Nonlinearities Be Bridged in Black Hole Ringdown?

Quadratic quasinormal modes (QQNMs) and Christodoulou memory effect are key nonlinear phenomena in gravitational wave physics. QQNMs characterize the near zone nonlinear response of a perturbed black hole, whereas the memory effect is a nonlinear remnant imprinted at null infinity by outgoing radiation. This naturally raises the question of whether and in what sense the two can be bridged. We show that they are related through bridge coefficients which depend primarily on remnant black hole parameters during ringdown. Future space-based gravitational wave detectors can probe this relation. These results provide a new avenue for testing gravity and a fresh perspective on the nonlinear regime of general relativity.

gr-qc↗

arXiv2026

Transition of vortex dipole dynamics in holographic superfluids

Using holographic duality, we reveal a transition in vortex dipole dynamics below a critical dipole size in strongly interacting superfluids, characterized by a significant suppression of mutual friction. In the bulk, this transition is triggered by a topological reconnection of vortex tubes, which disconnects the boundary vortices from the black hole horizon and forms a \textit{U-pipe}. Consequently, the post-transition evolution is governed by the contraction of the bulk \textit{U-pipe} rather than the mutual friction associated with the horizon, revealing a scale-dependent dissipation mechanism. We further show that this reconnection persists over a broad temperature range, even when the transition becomes unobservable at high temperatures. Our results provide a dissipation-based interpretation for the anomalous critical dipole scale observed in strongly interacting cold-atom experiments, and suggest the existence of distinct dissipative regimes in strongly interacting superfluids.

hep-th↗

arXiv2026

Explicit and covariant formula for thermodynamic volume in extended black hole thermodynamics

In extended black hole thermodynamics, the cosmological constant and other couplings are treated as thermodynamic variables, yielding the first law $\tildeδM = T\tildeδS+Ω\tildeδJ +\mathcal{V} \tildeδP+\cdots$, where $P\equiv -\fracΛ{8π}$. A long-standing conceptual gap in this framework is that, unlike $M$, $T$, $S$, $Ω$, and $J$, the thermodynamic volume $\mathcal{V} $ lacks a first-principles definition and can only be deduced from other thermodynamic quantities. This deficiency indicates that the underlying origin of $\mathcal{V} $ remains poorly understood. In this paper, we resolve this issue and provide an explicit, covariant formula for $\mathcal{V} $. We demonstrate that $\mathcal{V} $ (and the conjugate quantities of other couplings) universally decomposes into two contributions: one arising from the explicit coupling dependence of the Lagrangian, and the other from the response of the fundamental dynamical fields. This clarifies the physical meaning of the thermodynamic volume and places it on the same footing as other intrinsic thermodynamic quantities.

gr-qc↗

arXiv2026

Aero-World: Action-Conditioned Aerial Video Generation from Inertial Controls

Foundation video models produce visually impressive results, but their use in embodied AI remains limited because they are primarily trained on natural language rather than low-level control signals. This limitation is especially pronounced for aerial flight, where motion occurs in unconstrained 6-DoF space and small errors in ego-motion can produce large trajectory drift. Generating aerial videos that follow fine-grained inertial actions can support scalable training and evaluation of aerial agents by providing a controllable proxy for real-world or expensive simulation data. To address this problem, we propose \textbf{Aero-World}, a method for converting a pretrained image-to-video diffusion model into a controllable aerial video generator. Aero-World injects sequences of translational acceleration and angular velocity into a pretrained latent diffusion transformer through an action-token stream. A frozen latent-space Physics Probe, trained independently on real video--IMU pairs, provides differentiable inertial-consistency supervision during LoRA finetuning while avoiding computationally expensive video decoding. We further propose \textbf{AeroBench}, a benchmark for evaluating whether generated drone videos adhere to low-level action signals. AeroBench uses Action Alignment Score (AAS) to measure agreement with commanded inertial actions and Physical Consistency Rate (PCR) to measure temporal motion stability. On AeroBench, Aero-World improves mean AAS from 57.7 to 63.6 over action-only finetuning and gives a stronger quality-control trade-off than AirScape, with lower FVD (596.5 vs. 1058.6), higher SSIM (0.595 vs. 0.505), and higher Flow-IMU correlation (0.44 vs. 0.20). These results suggest that frozen Physics Probe supervision is a practical mechanism for adapting pretrained video generators toward more action-aligned aerial motion.

cs.CV↗

arXiv2026

Beyond Superficial Unlearning: Sharpness-Aware Robust Erasure of Hallucinations in Multimodal LLMs

Multimodal LLMs are powerful but prone to object hallucinations, which describe non-existent entities and harm reliability. While recent unlearning methods attempt to mitigate this, we identify a critical flaw: structural fragility. We empirically demonstrate that standard erasure achieves only superficial suppression, trapping the model in sharp minima where hallucinations catastrophically resurge after lightweight relearning. To ensure geometric stability, we propose SARE, which casts unlearning as a targeted min-max optimization problem and uses a Targeted-SAM mechanism to explicitly flatten the loss landscape around hallucinated concepts. By suppressing hallucinations under simulated worst-case parameter perturbations, our framework ensures robust removal stable against weight shifts. Extensive experiments demonstrate that SARE significantly outperforms baselines in erasure efficacy while preserving general generation quality. Crucially, it maintains persistent hallucination suppression against relearning and parameter updates, validating the effectiveness of geometric stabilization.

cs.LG↗

arXiv2026

AuralSAM2: Enabling SAM2 Hear Through Pyramid Audio-Visual Feature Prompting

Segment Anything Model 2 (SAM2) exhibits strong generalisation for promptable segmentation in video clips; however, its integration with the audio modality remains underexplored. Existing approaches either convert audio into visual prompts (e.g., boxes) via foundation models, or inject adapters into the image encoder for audio-visual fusion. Yet both directions fall short in human-in-the-loop scenarios due to limited prompt accuracy and increased inference overhead. In particular, these adapter-based methods often suffer from audio prompt dilution, where the signal gradually weakens as it propagates through the network. In this work, we propose AuralSAM2, which integrates audio into SAM2 while largely preserving its promptable segmentation capability. Its core module, AuralFuser, fuses audio and visual features to generate sparse and dense prompts. Guided by audio and built upon SAM2's feature pyramid, these prompts propagate auditory cues across visual layers, reinforcing cross-modal influence. To further align modalities, we introduce an audio-guided contrastive loss that emphasises auditory relevance in dominant visual features. Our method achieves notable accuracy gains on public benchmarks with only minimal impact on the interactive efficiency of promptable segmentation. Our code is available at https://github.com/yyliu01/AuralSAM2.

cs.CV↗

arXiv2026

Fermionic Love number of Reissner-Nordström black holes

The tidal deformation of compact objects, characterised by their Love numbers, provides insights into the internal structure of neutron stars and black holes. While static bosonic tidal Love numbers vanish for black holes in general relativity, it has been recently revealed that static fermionic tidal perturbations can induce non-zero Love numbers for Kerr black holes. In this paper, we investigate the response of the Reissner-Nordström black hole to the fermionic Weyl field. As a result, we find that the corresponding fermionic tidal Love numbers are also non-vanishing for the Reissner-Nordström black holes except for the extremal ones, which highlights the universal distinct behavior of the static fermionic tidal Love numbers compared to the bosonic counterparts.

gr-qc↗

arXiv2026

SafeSteer: A Decoding-level Defense Mechanism for Multimodal Large Language Models

Multimodal large language models (MLLMs) are gaining increasing attention. Due to the heterogeneity of their input features, they face significant challenges in terms of jailbreak defenses. Current defense methods rely on costly fine-tuning or inefficient post-hoc interventions, limiting their ability to address novel attacks and involving performance trade-offs. To address the above issues, we explore the inherent safety capabilities within MLLMs and quantify their intrinsic ability to discern harmfulness at decoding stage. We observe that 1) MLLMs can distinguish the harmful and harmless inputs during decoding process, 2) Image-based attacks are more stealthy. Based on these insights, we introduce SafeSteer, a decoding-level defense mechanism for MLLMs. Specifically, it includes a Decoding-Probe, a lightweight probe for detecting and correcting harmful output during decoding, which iteratively steers the decoding process toward safety. Furthermore, a modal semantic alignment vector is integrated to transfer the strong textual safety alignment to the vision modality. Experiments on multiple MLLMs demonstrate that SafeSterr can improve MLLMs' safety by up to 33.40\% without fine-tuning. Notably, it can maintain the effectiveness of MLLMs, ensuring a balance between their helpfulness and harmlessness.

cs.AI↗

arXiv2026

Two-Phonon Resonance Drives Multicomponent Mechanical Cat States

Using quadratic optomechanical coupling to prepare high-purity mechanical cat states is not feasible as its strength is several orders weaker than linear optomechanical coupling. Here, using only linear coupling in a multimode system, we achieve strong interaction between photons and two phonons, enabling the deterministic generation of high-purity multicomponent mechanical cats. Mediated by an auxiliary supermode, when other two optical supermodes satisfy the two-phonon resonance condition, the process whereby the annihilation of a high-frequency photon accompanied by the creation of a low-frequency photon and two phonons is strongly enhanced. Such resonant two-phonon process drives multiple rotations and interferences of mechanical coherent states, deterministically generating a multicomponent mechanical cat immune to both mechanical and optical losses. Our work provides an universal strategy for enhancing high-order phonon nonlinearities, paving the way for quantum state engineering, quantum precision measurement and fault-tolerant quantum computation.

quant-ph↗

arXiv2026

Skills-Coach: A Self-Evolving Skill Optimizer via Training-Free GRPO

We introduce Skills-Coach, a novel automated framework designed to significantly enhance the self-evolution of skills within Large Language Model (LLM)-based agents. Addressing the current fragmentation of the skill ecosystem, Skills-Coach explores the boundaries of skill capabilities, thereby facilitating the comprehensive competency coverage essential for intelligent applications. The framework comprises four core modules: a Diverse Task Generation Module that systematically creates a comprehensive test suite for various skills; a Lightweight Optimization Module dedicated to optimizing skill prompts and their corresponding code; a Comparative Execution Module facilitating the execution and evaluation of both original and optimized skills; and a Traceable Evaluation Module, which rigorously evaluates performance against specified criteria. Skills-Coach offers flexible execution options through its virtual and real modes. To validate its efficacy, we introduce Skill-X, a comprehensive benchmark dataset consisting of 48 diverse skills. Experimental results demonstrate that Skills-Coach achieves significant performance improvements in skill capability across a wide range of categories, highlighting its potential to advance the development of more robust and adaptable LLM-based agents.

cs.CL↗

arXiv2026

DPN-LE: Dual Personality Neuron Localization and Editing for Large Language Models

With the widespread adoption of large language models (LLMs), understanding their personality representation mechanisms has become critical. As a novel paradigm in Personality Editing, most existing methods employ neuron-editing to locate and modify LLM neurons, requiring changes to numerous neurons and leading to significant performance degradation. This raises a fundamental question: Are all modified neurons directly related to personality representation? In this work, we investigate and quantify this specificity through assessments of general capability impact and representation-level patterns. We find that: 1) Current methods can change personalities but reduce overall performance. 2) Neurons are multifunctional, connecting personality traits and general knowledge. 3) Opposing personality traits demonstrate distinctly mutually exclusive representation patterns. Motivated by these findings, we propose DPN-LE (Dual Personality Neuron Localization and Editing), which identifies personality-specific neurons by contrasting MLP activations between high-trait and low-trait samples. DPN-LE constructs layer-wise steering vectors and applies dual-criterion filtering based on Cohen's $d$ effect size and activation magnitude to isolate mutually exclusive neuron subsets. Sparse linear intervention on these neurons enables precise personality control at inference time. Using only 1,000 contrastive sample pairs per trait, DPN-LE intervenes on $\sim$0.5\% of neurons while achieving competitive personality control and substantially better capability preservation across reasoning tasks. Experiments on LLaMA-3-8B-Instruct and Qwen2.5-7B-Instruct demonstrate the effectiveness and generalizability of our approach.

cs.CL↗

arXiv2026

Exploring the Secondary Risks of Large Language Models

Ensuring the safety and alignment of Large Language Models is a significant challenge with their growing integration into critical applications and societal functions. While prior research has primarily focused on jailbreak attacks, less attention has been given to non-adversarial failures that subtly emerge during benign interactions. We introduce secondary risks a novel class of failure modes marked by harmful or misleading behaviors during benign prompts. Unlike adversarial attacks, these risks stem from imperfect generalization and often evade standard safety mechanisms. To enable systematic evaluation, we introduce two risk primitives verbose response and speculative advice that capture the core failure patterns. Building on these definitions, we propose SecLens, a black-box, multi-objective search framework that efficiently elicits secondary risk behaviors by optimizing task relevance, risk activation, and linguistic plausibility. To support reproducible evaluation, we release SecRiskBench, a benchmark dataset of 650 prompts covering eight diverse real-world risk categories. Experimental results from extensive evaluations on 16 popular models demonstrate that secondary risks are widespread, transferable across models, and modality independent, emphasizing the urgent need for enhanced safety mechanisms to address benign yet harmful LLM behaviors in real-world deployments.

cs.LG↗