arXiv subjects

Xin Zhang

Publications and source records attributed to Xin Zhang.

At least 127 records · Page 7Linked to original sources

arXiv2026

SHE: Trajectory-driven Safety Harness Evolution for LLM Agents

The safety of large language model (LLM) agents depends not only on model weights but also on the agent harness that manages context, memory, tools, permissions, and runtime control. Existing safety mechanisms often treat the harness as a fixed deployment artifact, limiting their ability to evolve with emerging risks. Moreover, coupled functions across harness components obscure safety responsibility attribution, making localized evolution difficult. We propose Safety Harness Evolution (SHE), a framework that learns evolving safe boundaries from rollout trajectories. SHE decomposes the harness into four artifacts with explicit safety responsibilities, including the System Prompt, Rule Bank, Safety Memory, and Tool Policy, defining clear functional boundaries for localized evolution. Based on this decomposition, SHE introduces an attribution-guided evolution loop that converts trajectory failures into structured diagnoses, learns artifact-specific boundary refinements, and selects evolved harnesses through safety-utility validation. Experiments on Agent-SafetyBench demonstrate that SHE effectively enhances safety through harness evolution, achieving a 3.1x ASR reduction compared with static SafeHarness, while also improving benign utility. The evolved harness further generalizes to unseen risks on the held-out AgentHarm benchmark and transfers across agent models without additional evolution.

cs.AI↗

arXiv2026

Conditional variational autoencoders for cosmological model discrimination and anomaly detection in cosmic microwave background power spectra

The cosmic microwave background power spectra are a primary window into the early universe. However, achieving interpretable compression and fast inference diagnostics under weak model assumptions remains challenging. We propose a parameter-conditioned variational autoencoder (CVAE) that aligns a data-driven latent representation with cosmological parameters while retaining an interface to likelihood-style diagnostic tests. The model achieves high directional reconstruction fidelity for the $D_\ell^{TT}$, $D_\ell^{EE}$, and $D_\ell^{TE}$ spectra in just 5 latent dimensions. It reconstructs spectra for several beyond-$Λ$CDM test cases, including controlled parameter extrapolations, and enables an amortized surrogate diagnostic that reduces one representative post-training MCMC run from $\sim$40 hours on CPU cores to $\sim$2 minutes on a GPU in this demonstration. The learned latent space shows a distributed, partially structured organization that mirrors known cosmological parameters and their degeneracies. It also provides representation-space discrimination diagnostics for distinguishing tested cosmological spectra from a fiducial reference. Overall, this physics-informed CVAE supports interpretable compression, rapid diagnostic exploration, and anomaly-sensitive representation learning beyond $Λ$CDM.

astro-ph.CO↗

arXiv2026

Interior geometry of black holes as a probe of first-order phase transition

Traditional diagnostics of black hole phase transitions rely on thermodynamic quantities defined at the event horizon or asymptotic boundary. Here, we demonstrate that the near-singularity geometry offers a sharp, independent probe of both first-order phase transitions and supercritical crossover. For scalarized AdS black holes exhibiting a first-order phase transition, the Kasner exponent $p_t$, which characterizes the approach to the singularity, undergoes a dramatic transformation. On one side of the transition, $p_t$ oscillates strongly with temperature, reflecting violent interior dynamics. On the other side, it becomes a smooth, monotonically varying function. These two distinct behaviors converge as the critical point is approached. Beyond the critical point, in the supercritical region, $p_t(T)$ develops a distinct extremum, defining a Kasner crossover line that is entirely independent of traditional thermodynamic (Widom line) or dynamic (Frenkel line) criteria. Our work establishes the near geometry of singularity of scalarized black hole as a novel class of diagnostics for phase transitions, revealing that a change in the macroscopic thermodynamic state fundamentally reshapes the deepest interior structure of spacetime.

gr-qc↗

arXiv2026

Is Personalized Modality Weighting Actually Personalized? A Controlled Audit of Per-User Weighting Claims in Multimodal Recommenders

Per-user modality weighting is deployed at billion-user scale in multimodal recommenders, through user modality-strength vectors, attention gates, meta-weight hypernetworks, and low-rank guided weights, each claiming a ranking gain from user-specific modality preference. Yet, to our knowledge, prior evaluations do not isolate a genuinely user-specific signal from a global modality weight plus model capacity. We audit this family with a two-contrast audit principle, reducing six implementations onto one shared collaborative backbone and measuring a utility gap (real-GM) against a single global modality weight and an identifiability gap (real-shuf) against an eval-time permutation of the user-weight binding. Across three independent short-video corpora, a single global weight already delivers nearly all of the content gain (+1.9/+3.6/+3.5pp over a no-modality baseline, p < .001). Making the weight per-user adds no consistent utility: no implementation wins on all corpora and metrics, and the few positive gaps are small (<=0.9pp) and flip. The shuffle control is necessary but not sufficient, since real-shuf reaches +128% of the content gain for heads that simultaneously lose to the global weight. We trace this dissociation to gates reading the shared collaborative embedding: decoupling the gate input collapses the inflated real-shuf to near zero while the utility conclusion stands. A monotone signal-implant dose-response (capture AUROC rising from 0.57 to 0.89 and from 0.64 to 1.00) verifies the harness would detect user-specific structure if present, and every finding replicates on a fourth, cross-domain e-commerce corpus. We propose reporting real-GM alongside real-shuf as a minimum evidentiary standard for personalization claims.

cs.IR↗

arXiv2026

Joint constraints on cosmic birefringence and early dark energy from ACT, Planck, DESI, and PantheonPlus

With the increasing number of high-precision astronomical observations, physical quantities that were previously inaccessible to accurate calculations, such as cosmic birefringence, have once again become a focal point of interest. Such phenomena induce a nonvanishing cross-correlation between the $E$- and $B$-mode polarizations of the cosmic microwave background (CMB), thereby providing a direct observational signature of parity violation. The Chern-Simons coupling between the scalar field in early dark energy (EDE) models and CMB photons is regarded as a plausible mechanism for generating cosmic birefringence. Recent data from the Atacama Cosmology Telescope (ACT) deliver $EB$ measurements at higher multipole moments than those previously achieved by {Planck}, while DESI and PantheonPlus datasets provide new and stringent constraints on the late-time expansion history. Using a joint analysis of {Planck}, DESI DR1, Pantheon+, and ACT data, we perform a full-parameter constraint on the cosmic birefringence effects induced by the EDE-CMB photon coupling. Our results favor a higher Hubble constant, $H_0 = 76.9^{+2.9}_{-2.5}\,\rm km\,s^{-1}\,Mpc^{-1}$, and a relatively large EDE fraction, $f_{\mathrm{EDE}} = 0.232^{+0.074}_{-0.047}$. By comparing the cosmological evolution of this model across different data combinations, we find that the ACT-$EB$ data combined with {Planck} + DESI + PantheonPlus provide good constraints to both early- and late-Universe observations.

astro-ph.CO↗

arXiv2026

STAGE: A Full-Screenplay Benchmark for Reasoning over Evolving Stories

Movie screenplays are a demanding testbed for long-form narrative understanding, as characters' goals, beliefs, knowledge, and relationships evolve continuously across scenes. However, existing benchmarks primarily evaluate isolated facts from the completed screenplay, leaving unassessed whether models can track the evolving state of characters as the story unfolds. We introduce STAGE, a benchmark over 151 English and Chinese full-length screenplays, built on a provenance-linked narrative backbone that recovers the state and epistemic access of each character at every point along its timeline. Three tasks derived from the backbone jointly probe whether models can maintain, explain, and act on evolving narrative state: Character Development Tracking updates a focal character's state between checkpoints, Cross-Scene Narrative Evolution Reasoning targets cross-scene state transitions, and In-Script Character Role-Playing requires responses bounded by the character's state and knowledge at a specified point. We identify three failure modes of current LLMs: silent forgetting under recursive state updating, limited cross-scene reasoning even when all relevant evidence is supplied, and a trade-off in role-playing where stylistic character fidelity and screenplay-grounded memory faithfulness are optimized by different memory-access strategies. STAGE thus provides a unified framework for diagnosing how current models fail to track, reason about, and enact story evolution.

cs.CL↗

arXiv2026

MoWorld: A Flash World Model

The future of World Models depends not only on scaling model capability, but also on scaling practicality and inference efficiency. High-frame-rate inference enables responsive perception, planning, and control in real-world autonomous systems. To this end, we present MoWorld, a cost-effective yet high-performance Flash World Model with an end-to-end framework spanning data generation, pre-training, distillation, and efficient inference, enabling up to 50 FPS real-time interaction with cinematic visual quality without the need of high-end GPUs. To enable large-scale real-world deployment, MoWorld jointly optimizes model capability and cost throughout the entire development pipeline. Specifically, unlike existing approaches that primarily rely on large-scale video corpora, MoWorld is built upon a scalable 3D-native data engine accumulated from our large-scale 3D vision and generative modeling pipeline, enabling the efficient construction of geometrically consistent training data across diverse real-world and synthetic environments. Based on this foundation, a curriculum cross-frame pre-training strategy for stable and scalable World Model learning, an efficient denoising-step distillation algorithm to reduce diffusion training cost, and a mixed-precision parallel inference framework for low-cost real-time deployment. MoWorld is the first real-time interactive World Model built on the Neural Processing Unit (NPU) and can achieves up to 50 FPS in such the devices, enabling practical and efficient deployment at scale. Comprehensive evaluations demonstrate that MoWorld achieves leading performance; notably, its average inference cost is only 30\%-50\% of that of existing World Models, providing a practical foundation for large-scale real-world applications of World Models. We also demonstrate diverse applications of MoWorld.

cs.CV↗

arXiv2026

Quantitative Particle Approximation for Controlled Nonlinear Filtering

We estimate convergence rates of value functions for particle approximations of a controlled nonlinear filtering problem. The state is a McKean--Vlasov diffusion on the flat torus, driven by hidden idiosyncratic noise and observed common noise. The filter---the conditional law of the state given the observations---serves as the state variable of the control problem, and the associated value function solves a second-order Hamilton--Jacobi--Bellman equation on the Wasserstein space. We approximate this problem by a centralized $N$-particle control problem with independent idiosyncratic noises and a common observation noise. The framework accommodates nonseparable rewards and controlled drifts. Since a single control is applied to the entire population, the Hamiltonian is defined by an optimization performed after integration over the population. Under smoothness of the data, uniform ellipticity, and regularity of this Hamiltonian, we establish uniform value-function error bounds of order $N^{-1/6}$ for $d=1$, $N^{-1/6}(\log N)^{1/3}$ for $d=2$, and $N^{-1/(3d)}$ for $d>2$. The proof combines a translation lift in the common-noise direction, Fourier--Wasserstein inf- and sup-convolutions, viscosity comparison, and particle derivative estimates uniform in $N$.

math.OC↗

arXiv2026

Gaussian curvature and Lyapunov exponent as probes of black hole phase transitions

First-order phase transitions of black holes have been extensively studied within thermodynamic frameworks, yet the corresponding evolution of spacetime geometric properties remains unclear. This paper establishes a purely differential geometric framework to probe such phase transitions by analyzing the curvature of unstable null orbits. Using the geodesic curvature of the null circular orbit in the optical metric to locate the light ring, we demonstrate that the corresponding Gaussian curvature $K$ serves as a direct geometric signature of the phase transition. During a first-order phase transition, the curve $K$ versus temperature $T$ exhibits a multivalued structure within the spinodal region, precisely mirroring the swallowtail behavior of the free energy. Numerical analysis of Hayward-Letelier-AdS black holes confirms the effectiveness of this geometric signature. Our work demonstrates that the intrinsic geometric quantities of spacetime encode the information of black hole phase transitions. These quantities serve as geometric probes of black hole phase transitions, while their discontinuity between the small and large black hole branches exhibits order parameter-like behavior. As an extension of this geometric probe, we also find that the Gaussian curvature exhibits a heat-capacity-like divergence at the second-order phase transition point. These results provide a purely geometric foundation for understanding the correspondence between thermodynamics and spacetime curvature in the null case.

gr-qc↗

arXiv2026

Nonequilibrium crossover in the supercritical region from quench dynamics

Distinguishing different subphases in the supercritical region is an important issue in statistical physics and condensed matter physics. Traditional approaches rely mainly on static thermodynamic response functions or equilibrium correlation functions, which are essentially limited to quasistatic processes. In this paper, we investigate the evolution behavior of a system after a rapid quench from the perspective of nonequilibrium dynamics within a holographic model. We find that, using the time at which inhomogeneous structures appear most rapidly, we can define a supercritical crossover curve based on the pure phase separation process. In addition, the uniform invasion phenomenon induced by topological defects persists in the supercritical region, and the invasion velocity exhibits a clear turning point as a function of the quench endpoint. This turning point can define another new nonequilibrium supercritical crossover line that simultaneously incorporates the effects of both symmetry breaking and phase separation. Unlike the classical Widom line or Frenkel line, these two new crossover lines contain both thermodynamic information and dynamical information, reflecting the dynamical nature of the supercritical region under nonequilibrium conditions. This work provides a novel nonequilibrium dynamical approach for characterizing supercritical subphases.

cond-mat.stat-mech↗

arXiv2026

Unifying topological, geometric, and complex classifications of black hole thermodynamics

Black hole thermodynamics has recently witnessed three distinct classification schemes: based on local geometric properties of the temperature function, global topological invariants, and Riemann surface foliations in the complex plane. We show that these schemes can be precisely mapped onto one another in the real domain via two dictionaries: one linking thermal stability to the monotonicity of the temperature curve, and the other connecting the number of black hole states to the foliation number of a Riemann surface. The number of extremal points of the temperature curve determines the classification in all three frameworks, tracing this unification to the critical point structure of the black hole solution space. As an illustration, several black holes demonstrate how counting extrema yields topological invariants and phase transition information. This unified framework simplifies black hole thermodynamic analysis and provides a foundation for exploring more complex black holes.

gr-qc↗

arXiv2026

PhysOmni: Physics-Grounded Multi-Object Scene Generation from a Single Image with Real-Time Interaction

Recent generative video models achieve impressive visual quality but remain constrained by limited physical consistency and controllability. Existing video generation methods provide minimal physical control, and single-image-to-3D conversion approaches often suffer from object interpenetration. Furthermore, physics-based scene-level 3D generation methods exhibit spatial misalignment, stylized artifacts, and inconsistencies with the input data, restricting their use in realistic interactive video synthesis. We propose PhysOmni, a training-free framework that converts a single image into a physically consistent and controllable video through holistic scene-level 3D reconstruction. By rep?resenting the full scene geometry in a unified spatial coordinate system, PhysOmni resolves object penetration and alignment ambiguity. Unlike prior methods, this formulation enables accurate scene?level multi-object interactions and introduces richer, complex control types for advanced mechanics?based manipulation. By decoupling simulation from rendering, PhysOmni bypasses latency-heavy priors, achieving real-time physical interaction previews paired while preserving photorealistic visual fidelity. Experimental results demonstrate that PhysOmni substantially outperforms prior methods in physical fidelity, spatial coherence, and controllability. Project Page: https://physomni.github.io/

cs.GR↗

arXiv2026

A Wireless Volumetric Metamaterial Resonator for Breast MRI

High-performance breast magnetic resonance imaging (MRI) remains limited by the availability, geometry, and cost of dedicated receive coils. In contrast, spine coils are widely installed in MR scanners globally but provide insufficient sensitivity for anterior breast tissue. Here, we report a wireless volumetric metamaterial resonator (VMR) that converts a standard built-in spine coil into a scalable breast MRI receiver platform without modifying the scanner hardware. The VMR consists of axially stacked coaxial resonant rings that support a co-rotating collective mode, producing a centrally concentrated RF magnetic field within the enclosed imaging volume. Leveraging distributed capacitance in the coaxial cable, gap-engineered geometric tuning, and PIN-diode-based self-detuning, the VMR enables size scalability, electric-field confinement, and transmit-receive compatibility, as validated through bench measurements and 3.0 T MRI experiments. Three VMR prototypes spanning small, medium, and large breast-equivalent volumes were designed and evaluated using electromagnetic simulations, bench measurements, and phantom MRI experiments. Compared with a commercial 16-channel breast coil, the VMR-augmented spine coil achieved more than 4-fold higher central SNR and reduced spatial non-uniformity from approximately 70% to 20% within the target imaging region. Under SENSE parallel imaging at acceleration factors of 2-4, the VMR-augmented spine coil maintained an average 4-fold higher central SNR than the commercial coil for three sizes. These results establish the VMR as a passive, scalable hardware strategy for improving breast MRI performance using existing clinical scanner infrastructure.

physics.med-ph↗

arXiv2026

Self-Consistent Parker Bound on Magnetic Monopoles

Magnetic monopoles arise generically in unified theories and offer a natural explanation of charge quantization. Beyond collider searches and cosmic-ray experiments, their flux is constrained by Parker-type bounds requiring galactic magnetic fields to survive monopole energy extraction. We formulate a self-consistent Parker bound anchored in the lowest eigenmode of the galactic mean-field dynamo and convert the resulting limit to the present-day flux. Small-scale turbulent fields both seed this eigenmode and set the monopole velocity via stochastic acceleration before energy extraction from the coherent field. These unavoidable effects substantially modify the standard extended Parker bound at low and intermediate masses, yielding flux limits robust to primordial magnetic fields (PMFs); PMFs strong enough to alter these limits lie in regimes constrained by Ly$α$ data or testable by 21-cm observations and other cosmological probes.

hep-ph↗

arXiv2026

Electron-Stimulated Desorption of D Atoms from Gibbsite (Al(OD)3) and D2O Ice: Energy and Temperature Dependence of Translational Energy Distributions

The electron-stimulated desorption (ESD) of neutral D atoms from gibbsite (\ce{Al(OD)3}) nanoplatelets and amorphous \ce{D2O} ice has been investigated using $2+1$ resonance-enhanced multiphoton ionization (REMPI) time-of-flight mass spectroscopy in a high vacuum chamber at temperatures 15 and 300\,K. Electron irradiation at 540, 250, and 150\,eV produces similar translational energy distributions at $\sim$300\,K, with a dominant intermediate-temperature component ($T \sim 1500$--$2100$\,K). Cooling to 15\,K suppresses the D atom yield by approximately 50\% and removes the lowest-temperature (slowest) component. This decrease in yield is consistent with diminished hole mobility and restricted diffusion at cryogenic temperatures. Under identical conditions, \ce{D2O} amorphous solid water ice films produce approximately 20 times greater D atom signal than bare gibbsite, with significantly hotter translational distributions, reflecting the higher deuterium surface density and distinct bonding environments of bulk ice relative to the terminal hydroxyl groups on gibbsite. These results identify hole transport to terminal hydroxyl sites as the rate-limiting step for nonthermal D atom production and provide a mechanistic framework for understanding atomic hydrogen release from aluminum hydroxide phases relevant to radioactive waste storage at the Hanford Site.

physics.chem-ph↗

arXiv2026

Cylindrical Projections of Occupied Diffusions

Occupied diffusions offer a Markovian framework for path-dependent dynamics by lifting the state space with a flow of occupation measures. Because this additional feature is infinite-dimensional, the simulation of these processes remains computationally intractable. We address this by introducing \textit{cylindrical projections}, which approximate the occupation flow via a finite-dimensional system. We establish the strong convergence of this approximation to the initial process and derive corresponding convergence rates. The method is validated through Euler--Maruyama simulations of self-interacting diffusions and an application to the Local Occupied Volatility (LOV) model in finance. Finally, we provide a weak error analysis and explore its consequences for Monte Carlo methods and derivatives pricing.

math.NA↗

arXiv2026

An MLIR-Based Compilation Method for Large Language Models

Large Language Models (LLMs) have become the dominant workload on modern AI accelerators, yet deploying them on specialized hardware still faces two core challenges: how to import a trained model into a compiler-friendly intermediate representation, and how to efficiently schedule the autoregressive inference loop under limited on-chip memory. This paper presents an MLIR (Multi-Level Intermediate Representation) based compilation method for large language models, illustrated using two dialects of operators, TopOp and TpuOp. TopOp serves as a high-level graph dialect that is independent of both the source framework and the target chip, and is responsible for expressing model semantics; TpuOp serves as the target hardware dialect, carrying chip-related decisions such as quantization, layer groups, and memory layout. A model is first represented as TopOp, then lowered layer by layer to TpuOp, and finally a deployable binary is generated. In addition, each Transformer layer is split into three stages for static compilation: prefill, prefill_kv (prefill with historical key-value cache), and decode, so as to accommodate the different computational characteristics of prompt-parallel processing and per-token generation. The method has been implemented in the TPU-MLIR compiler {https://github.com/sophgo/tpu-mlir} and the LLM-TPU deployment project {https://github.com/sophgo/LLM-TPU}, supporting a variety of generative models including the Qwen, Llama, InternVL, and MiniCPM-V series, as well as multiple quantization and deployment forms such as GPTQ, AWQ, and AutoRound.

cs.CL↗

arXiv2026

Physics Closure Matters for Machine Olfaction: A Maxwell-Stefan Graph Solver for Identifiable Dynamic Gas Unmixing

Machine olfaction for gas unmixing is an underconstrained inverse problem in which gas compositions must be inferred from low-dimensional, delayed, and entangled sensor responses produced by interacting chemical transport, surface adsorption, and sensor transduction. One of the key obstacles is physics closure misspecification, where a neural network is designed to fit sensor traces rather than infer a physically closed olfactory process. In this work, we formulate gas unmixing as a multi-physics-constrained inverse problem governed by Maxwell--Stefan multicomponent transport PDEs, competitive adsorption ODEs, and nonlinear sensor transduction ODEs. Directly solving such a high-dimensional coupled system is computationally expensive and often numerically unstable. To this end, we propose UnMixNet, a physics-closed graph neural solver that embeds this multi-physics forward process into end-to-end gas unmixing. UnMixNet discretizes Maxwell--Stefan cross-diffusion on spatial graphs and formulates the multicomponent flux on each edge. This design enables local, differentiable, and flux-conservative inference for multicomponent cross-diffusion. Evaluations on SmellNet show improved single-odor recognition, seen-mixture unmixing, and unseen-mixture generalization. In addition, an external validation on UCI Dynamic Gas Mixtures shows that the inferred concentration process agrees with ground truth concentration set points under dynamic transitions. Process-consistency diagnostics further show that the proposed model learns transferable dynamic physical fingerprints that better satisfies transport, conservation, adsorption, and readout closure.

cs.CV↗