arXiv subjects

Peng Cheng

Publications and source records attributed to Peng Cheng.

At least 19 recordsLinked to original sources

arXiv2026

Identifiability Limits of Reynolds-Stress Attribution in Localized Nonorthogonal Flow Representations

Reynolds stress is assembled from correlations among velocity fluctuations; assigning it to named vortical structures additionally requires their velocity fields to be separable. We examine this requirement in homogeneous isotropic turbulence from the Johns Hopkins Turbulence Databases (JHTDB) using a frozen solenoidal expansion of curl-lifted anisotropic Gaussian vorticity functions. The expansion retains locally addressable blob-, sheet-, and tube-like geometry, while its fixed Biot-Savart map carries coefficient covariance exactly into the stress of the represented Gaussian velocity. Principal-angle geometry and self-cross covariance then connect overlap between group velocity spans to stress assembly. At a principal cosine of 0.95, 55% of the sheet-associated directions are also closely represented by the tube-like family, and products within the shared directions carry 99.6% of the covariance cancellation. Removing between-family covariance increases the direct numerical simulation (DNS)-referenced stress error by factors of 53.6 to 87. Orthogonal Fourier and proper orthogonal decomposition (POD) regroupings of the same velocity remove this cancellation, showing that it follows from the local nonorthogonal organization rather than from the represented field alone. The morphology labels remain representation-dependent, but their physical consequence is direct: Reynolds stress in this local description is carried by coupled velocity fluctuations, not by additive component energies. Localized nonorthogonal representations can therefore reveal how turbulent momentum flux is assembled, provided that span overlap and cross covariance are retained.

physics.flu-dyn↗

arXiv2026

Self-Supervised Passive WiFi CSI-Based Motion Tracking with Physics-Informed Deep SSM

Inferring target position from WiFi channel state information (CSI) is challenging due to the noisy and heterogeneous nature of CSI-derived observations. Measurements such as angle, delay, and Doppler provide only imperfect information about target position and velocity, particularly in dynamic multipath environments where observation reliability varies. Motivated by this challenge, this paper proposes a physics-informed Bayesian state space framework for passive CSI-based tracking. We introduce a context-conditioned deep state space model (DSSM) that captures temporal motion and environmental dynamics, while adaptively modeling time-varying observation uncertainties. By incorporating geometric constraints, this framework enables heterogeneous CSI measurement modalities to jointly regularize the consistency and confidence of the tracking process. Experimental results demonstrate that the proposed framework yields smoother trajectory estimates with reduced temporal fluctuations and outperforms existing methods in tracking accuracy.

eess.SP↗

arXiv2026

Modified anisotropy and distinct phase diagrams in the nearly isotropic $S=1$ triangular-lattice antiferromagnet $\mathrm{Rb_2Ni(SeO_3)_2}$

The triangular-lattice antiferromagnet serves as a benchmark platform for realizing exotic quantum phases, in which magnetic anisotropy plays an essential role. In this work, we report modified magnetic anisotropy and phase diagrams in the nearly isotropic $S=1$ triangular-lattice antiferromagnet $\mathrm{Rb_2Ni(SeO_3)_2}$, compared with those of its isostructural analog $\mathrm{K_2Ni(SeO_3)_2}$. The magnetic susceptibility of $\mathrm{Rb_2Ni(SeO_3)_2}$ exhibits an easy-axis anisotropy with $g_c > g_{ab}$, opposite to the easy-plane anisotropy observed in $\mathrm{K_2Ni(SeO_3)_2}$. The single Néel transition observed at zero field splits into two transitions at around 10~T, and a magnetization plateau is observed between 21~T and 37~T for both out-of-plane ($H\parallel c$) and in-plane ($H\parallel ab$) field directions in $\mathrm{Rb_2Ni(SeO_3)_2}$. We construct magnetic phase diagrams that are distinct between $\mathrm{Rb_2Ni(SeO_3)_2}$ and $\mathrm{K_2Ni(SeO_3)_2}$. These findings highlight the sensitivity of field-induced magnetic phase behavior to anisotropy in nearly isotropic triangular-lattice antiferromagnets.

cond-mat.str-el↗

arXiv2026

Training Matryoshka Mixture-of-Experts for Elastic Inference-Time Expert Utilization

Mixture-of-Experts (MoE) has emerged as a promising paradigm for efficiently scaling large language models without a proportional increase in computational cost. However, the standard training strategy of Top-K router prevents MoE models from realizing their full potential for elastic inference. When the number of activated experts is altered at inference time, these models exhibit precipitous performance degradation. In this work, we introduce Matryoshka MoE (M-MoE), a training framework that instills a coarse-to-fine structure directly into the expert ensemble. By systematically varying the number of activated experts during training, M-MoE compels the model to learn a meaningful ranking: top-ranked experts collaborate to provide essential, coarse-grained capabilities, while subsequent experts add progressively finer-grained detail. We explore this principle at multiple granularities, identifying a layer-wise randomization strategy as the most effective. Our experiments demonstrate that a single M-MoE model achieves remarkable elasticity, with its performance at various expert counts closely matching that of an entire suite of specialist models, but at only a fraction of the total training cost. This flexibility not only unlocks elastic inference but also enables optimizing performance by allocating different computational budgets to different model layers. Our work paves the way for more practical and adaptable deployments of large-scale MoE models.

cs.CL↗

arXiv2026

Higher-Order Action Supervision Makes A Strong Policy Class

Modern data-driven decision-making methods, such as imitation learning (IL) and reinforcement learning (RL), have achieved great success in solving many complex tasks. However, these methods often suffer from serious control instability and robustness issues when applied in real-world applications such as robotics and autonomous driving, posing notable challenges for their practical deployment. We argue that this instability issue stems largely from their limitations in solely supervising and optimizing zeroth-order actions (i.e., the action labels), failing to account for higher-order action dynamics and temporal consistency. In this paper, we show that simultaneously supervising both zeroth- and first-order actions can dramatically enhance policies' performance and control robustness. To achieve this, we introduce a novel and elegant loss scheme supported by formal theoretical guarantees that can equip any off-the-shelf policy model (e.g., deterministic, stochastic, or flow policies) with the capability for higher-order action supervision, without requiring any structural modifications. Moreover, our proposed method can serve as a lightweight plug-and-play module that seamlessly integrates with a broad spectrum of existing offline RL frameworks. Extensive evaluations on OGBench and D4RL demonstrate that our approach yields substantial performance and robustness improvements across a wide range of continuous control environments. Notably, our method can also enhance policies' out-of-distribution (OOD) generalization capability in the challenging low-data regime, making it an ideal tool in tackling many real-world control problems.

cs.RO↗

arXiv2026

Rethinking Latency Denial-of-Service: Attacking the LLM Serving Framework, Not the Model

LLM inference is inherently expensive, even a modest slowdown can translate into substantial operating costs and severe availability risks. Recently, a growing body of research known as latency attacks focuses on crafting inputs to trigger worst-case output lengths. However, we report a contrary finding that these algorithmic-level latency attacks are largely ineffective against modern LLM serving systems. We reveal that system-level optimization such as continuous batching provides a logical isolation to mitigate contagious latency impact on co-located users. Thus, in this paper, we shift our focus from the algorithm to the system layer, and introduce a new Fill and Squeeze attack strategy targeting the state transition of the scheduler. ``Fill'' first exhausts the global KV cache to induce Head-of-Line blocking, while ``Squeeze'' forces the system into repetitive preemption. By manipulating output lengths using different attack prompts, and leveraging side-channel probing of memory status, we demonstrate that the attack can succeed in a practical black-box setting with much less cost. Extensive evaluations on vLLM indicate up to $75-742\times$ TTFT degradation relative to benign baselines and $1.5-4\times$ average slowdown on Time Per Output Token compared to existing attacks with 30-40% lower attack cost. Code: https://github.com/Phil-Fan/FS-attack

cs.CR↗

arXiv2026

MixFake: Benchmarking and Enhancing Audio Deepfake Detection in Diverse Real-world Mixed Audio

Speech deepfake detection has achieved remarkable success in clean environments but faces significant challenges in complex, real-world scenarios where speech is often mixed with background music or noise. Current state-of-the-art methods rely on semantic features from self-supervised learning (SSL) models, which often fail when processing non-speech or mixed-source audio. In this paper, we first introduce MixFake, a large-scale benchmark dataset designed to simulate diverse acoustic environments with varying SNR levels and mixed authenticity components. To address the "semantic-centric" limitation, we propose a Multi-stream Prompt Tuning framework that injects signal-level priors into SSL backbones. By integrating base, frequency, and texture streams through deep prompt injection, our model effectively captures acoustic artifacts. Experimental results demonstrate that our method significantly outperforms existing baselines, achieving a 0.95% EER in foreground detection and a substantial 7.72% absolute improvement in complex background detection tasks. Our dataset and code are available at https://github.com/saltfish233/MixFake.

cs.SD↗

arXiv2026

Task-Aware Joint Pruning and Distillation for Efficient Audio Deepfake Detection

Advances in speech synthesis have made deepfake speeches increasingly convincing, posing growing threats to security. While self-supervised learning (SSL) based detectors achieve state-of-the-art performance, their computational demands (typically 300M+ parameters) prevent deployment on resource-constrained devices. Existing compression methods, designed mainly for content-centric tasks, struggle to maintain competitive performance when directly adapted to deepfake detection. We propose a Task-Aware Joint Pruning and Distillation framework that combines cross-domain knowledge distillation with movement-guided structured pruning to transfer forgery-discriminative knowledge and preserve critical structures under aggressive compression. Our framework reduces the model to 31.9M parameters with 6.3$\times$ FLOPs reduction, with an average performance drop of only 1.30\% across multiple datasets compared to the uncompressed baseline, demonstrating strong potential for on-device deployment.

cs.SD↗

arXiv2026

Near-horizon soft theorems from local Ward identities

We develop a unified formulation of leading soft photon theorems for nondegenerate static spherical horizons. By fixing the photon mode normalization from the horizon symplectic form, we obtain a common angular soft factor whose sign follows from the boundary orientation. We determine the geometric coefficients in the local angular and momentum representations for a chosen radial normalization. The charge balance includes horizon endpoint data and contributions from other boundaries. This formulation brings black hole and cosmological horizons into a common description. Applying the same construction to Yang-Mills fields yields the corresponding tree-level soft gluon theorem. These results identify the common soft structure underlying different horizon geometries.

hep-th↗

arXiv2026

Saddle-Order Universality in Ensemble-Averaged Black Hole Thermodynamics

We establish a saddle-order proposition for the ensemble-averaged generalized free energy in reduced Euclidean black hole ensembles. Under regularity and stability assumptions, the leading fluctuation contribution is determined by the order of the first stabilizing term around the dominant saddle. An ordinary critical point is governed by a quartic soft mode, giving a quarter-temperature contribution instead of the familiar half-temperature Gaussian result. We verify this analytically for four-dimensional RN-AdS and Kerr-AdS black holes and map their thermodynamic controls to a common quartic normal form, whose scaling function connects the critical and Gaussian regimes along the zero-field one-well direction. The proposition also extends to coupled collective modes with weighted-homogeneous leading potentials. These results provide a unified description of reduced-ensemble fluctuations through black hole criticality.

hep-th↗

arXiv2026

RoofLang: Enabling AI-Driven Architecting of LLM Inference Systems

AI is beginning to make substantive contributions to LLM inference optimization. Existing AI optimizations are predominantly profiling-based. Profiling-bound feedback confines the search to the capabilities and performance of an existing software stack, preventing a fundamentally better architecture of LLM inference systems from being identified. To enable the AI-driven LLM inference system architecting loop, we argue that a general workload representation, a verifiable mutation space, and an implementation-independent evaluator are required. We present the RoofLang domain-specific language (DSL) that provides these features. In our evaluation, RoofLang reveals that DeepSeek V4-series models could achieve 3.5-39.5$\times$ higher peak decode throughput than other representative models. This gap is disproportionate to their total parameter counts and arises largely from compact KV-cache designs that support larger batches and reduce memory traffic. A persistent optimizer agent further discovered several new architectures that improved both throughput and interactivity of DeepSeek V4 Pro on NVIDIA B300 by 6.23-50.1%.

cs.DC↗

arXiv2026

Exploiting Residual Reachability for Cross-Model Migration of Graph-Based Indexes in Approximate Nearest Neighbor Search

Approximate nearest neighbor search (ANNS) underpins large-scale vector retrieval in search, recommendation, and retrieval-augmented generation. Graph-based indexes have demonstrated state-of-the-art search performance for ANNS. They connect each corpus vector to a small set of nearby or navigationally useful vertices and answer queries by traversing the resulting graph. Because these edges are selected using construction-time distances, the graph index is tied to the embedding model. Re-encoding a corpus with a new model may change distances and neighborhoods of the vectors. Reconstructing the graph for the new embedding vectors incurs substantial construction cost and delays deployment. When the embedding model changes, we observe a phenomenon in the old graph index that we call residual reachability. Specifically, although derived from different models, the vectors describe the same underlying objects and often retain part of their similarity structure. These shared relations are reflected in the connectivity of the old graph index, leaving many exact new-model neighbors reachable within a few hops in the old graph index. Motivated by this observation, we develop an index-migration approach that utilize the residual reachability in the old graph index to faster construct the new graph index for the new embedding vectors. Our method, Drift-Guided Migration (DGM), provides two migration paths. DGM-Local performs parallel shallow expansion over the inherited graph index and screens second-hop candidates with packed position sign codes before exact evaluation. DGM-Search uses hop-bounded beam traversal to explore beyond shallow expansion. Across eight text and image migrations, our DGM methods can achieve up to 17.43 times speedup on constructing the new graph index than the fastest degree-matched reconstruction method while keeping competitive recalls.

cs.DB↗

arXiv2026

VLM-MPPI: Grounding Natural Language in Behaviorally Diverse Trajectories for Aerial Navigation

We present a hierarchical UAV navigation framework that aligns natural-language intent with dynamically feasible flight behaviors in cluttered indoor environments. To bridge the gap between abstract semantics and low-level control, we employ a parallelized ensemble of six behavior-conditioned Model Predictive Path Integral (MPPI) planners. Crucially, by designing mode-specific guiding costs and sampling biases, we induce distinct trajectory modes that converge to unique behavioral means, yielding a compact set of intentionally diverse candidates rather than mere stochastic variations. We project these 3D candidates onto the onboard first-person-view RGB stream, turning language grounding into a visual action selection problem. A pretrained vision--language model (VLM) asynchronously selects the candidate index given the overlaid FPV image and a natural-language prompt, while MPPI replans at 20Hz and a PID-based low-level controller tracks the selected trajectory. We implement the full pipeline in NVIDIA Isaac Sim and on a real-world quadrotor platform equipped with LiDAR and RGB sensing. Experiments in both simulation and real-world flights show semantically meaningful behavior diversity, robust language alignment despite VLM latency, and safe, repeatable flight across all modes, achieving 100% task success in our evaluated scenarios.

cs.RO↗

arXiv2026

EnhanceGraph: A Continuously Enhanced Graph-based Index for High-dimensional Approximate Nearest Neighbor Search

Recently, Approximate Nearest Neighbor Search in high-dimensional vector spaces has garnered considerable attention due to the rapid advancement of deep learning techniques. We observed that a substantial amount of search and construction logs are generated throughout the lifespan of a graph-based index. However, these two types of valuable logs are not fully exploited due to the static nature of existing indexes. We present the EnhanceGraph framework, which integrates two types of logs into a novel structure called a conjugate graph. The conjugate graph is then used to improve search quality. Through theoretical analyses and observations of the limitations of graph-based indexes, we propose several optimization methods. For the search logs, the conjugate graph stores the edges from local optima to global optima to enhance routing to the nearest neighbor. For the construction logs, the conjugate graph stores the pruned edges from the proximity graph to enhance retrieving of k nearest neighbors. Our experimental results on several public and real-world industrial datasets show that EnhanceGraph significantly improves search accuracy with the greatest improvement on recall from 41.74% to 93.42%, but does not sacrifices search efficiency. In addition, our EnhanceGraph algorithm has been integrated into Ant Group's open-source vector library, VSAG.

cs.DB↗

arXiv2026

FGIM: a Fast Graph-based Indexes Merging Framework for Approximate Nearest Neighbor Search

As the state-of-the-art methods for high-dimensional data retrieval, Approximate Nearest Neighbor Search (ANNS) approaches with graph-based indexes have attracted increasing attention and play a crucial role in many real-world applications, e.g., retrieval-augmented generation (RAG) and recommendation systems. Unlike the extensive works focused on designing efficient graph-based ANNS methods, this paper delves into merging multiple existing graph-based indexes into a single one, which is also crucial in many real-world scenarios (e.g., cluster consolidation in distributed systems and read-write contention in real-time vector databases). We propose a Fast Graph-based Indexes Merging (FGIM) framework with three core techniques: (1) Proximity Graphs (PGs) to k Nearest Neighbor Graph (k-NNG) transformation used to extract potential candidate neighbors from input graph-based indexes through cross-querying, (2) k-NNG refinement designed to identify overlooked high-quality neighbors and maintain graph connectivity, and (3) k-NNG to PG transformation aimed at improving graph navigability and enhancing search performance. Then, we integrate our FGIM framework with the state-of-the-art ANNS method, HNSW, and other existing mainstream graph-based methods to demonstrate its generality and merging efficiency. Extensive experiments on six real-world datasets show that our FGIM framework is applicable to various mainstream graph-based ANNS methods, achieves up to 3.5X speedup over HNSW's incremental construction and an average of 7.9X speedup for methods without incremental support, while maintaining comparable or superior search performance.

cs.DB↗

arXiv2026

Universal Lichnerowicz Lifting of Near-Horizon Soft Modes

A recurring universality appears in the low-temperature quantum thermodynamics of near-extremal black holes, where distinct parent geometries often lead to the same logarithmic temperature dependence at one loop. In this work, we study the Lichnerowicz spectral origin of this infrared universality and understand why the relevant spectral data become insensitive to the details of the parent geometry. For extremal near-horizon geometries containing a two-dimensional maximally symmetric throat, we construct the normalizable transverse-traceless tensor zero modes associated with near-horizon reparametrizations. Turning on a small temperature lifts these zero modes through the first-order deformation of the Lichnerowicz operator. Although the local matrix element depends on detailed parent-geometry data, these data cancel after projection onto normalized tensor modes, leaving the universal result. For static spherically symmetric backgrounds, the resulting eigenvalue shift is proportional to the Fourier mode number and temperature. For rotating backgrounds, the same structure persists in the near-horizon reparametrization tensor sector, while angular warp factors only modify the overall projection factor. We fix the normalized coefficient explicitly, showing how radial integration and bulk normalization remove the near-horizon Taylor coefficients. We further show that this lifted bulk spectrum is the Lichnerowicz realization of the Schwarzian soft sector. Thus, the universal one-loop temperature dependence in the tensor sector is traced to an infrared bulk-boundary matching between near-horizon tensor zero modes and boundary reparametrization dynamics.

hep-th↗

arXiv2026

Expert-Space Exploration in MoE Reinforcement Learning

Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixture-of-Experts (MoE) models have primarily focused on improving optimization stability and training efficiency, while treating the expert selection as a fixed component. Since routing determines the sparse computation paths that induce output distributions, expert selection offers an additional source of rollout diversity. Through empirical analysis, we find that perturbing expert routing effectively alters model output and increases rollout diversity, which is similar to increasing the decoding temperature. However, direct perturbation can activate unsuitable experts and substantially degrade rollout quality. Motivated by these observations, we introduce Expert-Space Exploration Reinforcement Learning (ESRL), an architecture-aware framework that explicitly explores the expert-routing space of MoE models. ESRL preserves high-confidence experts as anchors, and restricts stochastic routing to a plausible candidate pool, thereby retaining reliable computation paths. The perturbation strength is further adapted according to router entropy to avoid over-perturbation. To mitigate the routing mismatch introduced by perturbation, ESRL records the expert paths used during rollout and replays them during policy optimization. Experiments demonstrate that ESRL achieves the best performance across MoE backbones with top-K, top-1, and shared-expert routing, as well as across mathematics, science, and code tasks without additional sampling or computational cost. Specifically, ESRL on Qwen3-30B-A3B achieves the best among all compared methods, improving average Pass@1 and Pass@8 over GRPO by 3.2 and 4.5 percentage points, respectively. Further analyses of expert utilization and training dynamics provide insights into how exploiting MoE-specific routing structure benefits RL training.

cs.CL↗

arXiv2026

Vacancy Order and Physical Properties of a Ternary Compound Fe0.68Pd0.80Te with an α-Fe1+xTe-Type Structure

We report the identification and characterization of a new compound Fe0.68Pd0.80Te with an α-Fe1+xTe prototype structure. Different from the Fe-square net and minor occupancy of interstitial Fe-sites in Fe1+xTe, Fe0.68Pd0.80Te is featured by a Pdsquare net and near 68% occupancy of the corresponding interstitial Fe-sites. Furthermore, noncontact atomic force microscopy and X-ray diffraction provide evidence for the existence of a 3*3*3 Pd-vacancy order in this layered material. A spin-glass ground state below Tg 40 K is identified via magnetic characterization. Electrical transport measurements show that Fe0.68Pd0.80Te is a semiconductor with a very small band gap below 10 meV. It has weak negative magnetoresistance and holelike charge carriers below room temperature. Our results demonstrate its potentials for further exploring various quantum phenomena.

cond-mat.mtrl-sci↗