arXiv subjects

Rong Chen

Publications and source records attributed to Rong Chen.

At least 19 recordsLinked to original sources

arXiv2026

UMLoad: Fast On-Demand Model Loading on Unified Memory Architecture

On-demand model loading dominates response latency for interactive on-device AI, yet unified memory architectures (UMAs), now common on edge devices, should make it cheap by placing weights in memory the GPU already sees. However, frameworks still use the loading flow built for discrete GPUs, leaving significant inefficiencies. This paper presents UMLoad, a system-level model-loading accelerator that decouples a framework's logical allocations from their physical page backing. UMLoad uses on-demand reconstruction to eliminate the weight copy while preserving copy independence, and GPU-assisted page zeroing to accelerate runtime allocation while preserving its security guarantee. Implemented on Intel, AMD, and NVIDIA UMA platforms and evaluated with llama.cpp, PyTorch, and vLLM across six models, UMLoad reduces average loading time by 36.6%-71.9% over vanilla and 19.9%-62.7% over pipelined loaders, and cuts mean LLM-routing time to first token (TTFT) by 51.3%-59.3% and model-backed tool completion latency by 10.2%-31.6%.

cs.OS↗

arXiv2026

Apto: Self-Adaptive CPU-GPU Hybrid MoE Inference on Shared PCs

CPU/GPU hybrid inference enables Mixture-of-Experts (MoE) models to run on memory-constrained PCs by keeping all expert weights in host memory and executing activated experts across both processors. Existing hybrid systems, however, assume exclusive resource access and determine an upfront CPU/GPU partition, so CPU contention from colocated applications can severely imbalance the two execution paths. This paper presents Apto, a self-adaptive hybrid inference system that applies dynamic scheduling to expert execution. The CPU and GPU executors claim experts from opposite ends of a dual-ended queue as they progress, allowing the CPU/GPU work split to adapt to their execution progress without predicting contention. Apto further uses affinity-aware scheduling to expose each executor's preferred experts at its queue end and cancellable tail racing to mitigate final-expert stragglers. Compared with state-of-the-art GPU-only and hybrid baselines, Apto achieves average prefill and decode speedups of 2.96$\times$ and 1.52$\times$ under standalone execution, and 5.97$\times$ and 3.45$\times$ under concurrent CPU workloads. It also accelerates real-world coding-agent applications by 2.02$\times$ on average.

cs.DC↗

arXiv2026

Towards Self-Evolving Narrative Systems: Real-Time Adaptation and Long-Term Optimization with LLMs

Traditional interactive narrative systems offer strong authorial control and structural coherence but often restrict player agency. In contrast, LLM-driven narratives enable open-ended generation and greater flexibility, yet frequently suffer from inconsistency and diminished narrative control. To navigate this trade-off, we propose a hybrid narrative framework that dynamically balances pre-authored content with real-time LLM generation via semantic similarity alignment. Additionally, sustaining narrative quality over extended gameplay remains a core challenge in LLM-based storytelling. To address this, we introduce a self-evolving, feedback-driven optimization framework that leverages in-context learning to iteratively refine prompts, eliminating the need for manual prompt engineering. Through a case study and participant-based evaluation, we show that our framework enhances narrative coherence while preserving player agency, effectively aligning generated content with both authorial intent and evolving player behavior.

cs.CL↗

arXiv2026

PhOSv1: Efficient Scaling of Model Sleeping on a Single Server for Autoscaling with GPU-Context Deduplication

Autoscaling is a key mechanism in model-as-a-service systems, where fast scale-up is essential because the scale-up time directly determines how long a model service remains overloaded, which in turn governs serving quality. The state-of-the-art approach, model sleeping, adopted by mainstream serving engines including vLLM and SGLang, achieves near-zero-cost scale-up by pausing pre-spawned engines on the GPU. However, sleeping is not a free lunch: each sleeping model must keep its GPU context resident, and this cost grows linearly with the number of models, degrading serving performance. Because providers must support many models, they face a dilemma between fast scale-up for all models and high service performance. In this paper, we show that a single server can keep many models asleep without paying this GPU context tax. The key observation is that the contexts of sleeping models are largely duplicated: the driver (CUDA) context can be shared across models, and their GPU kernel binaries overlap heavily due to common operators and libraries. PhOSv1 deduplicates this context by transparently delegating each engine's GPU context allocation to a central daemon through API remoting; therefore, a single server can keep far more models asleep than current engines allow. Our evaluation shows that PhOSv1 matches the scale-up time of vanilla vLLM sleep mode while substantially reducing the GPU memory footprint, improving end-to-end multi-model serving performance.

cs.DC↗

arXiv2026

Theta Relations

Motivated by the classical quartic and cubic theta identities, we establish a product formula for Siegel--Jacobi theta series associated with rationally isometric positive-definite quadratic spaces. Our main theorem expresses a product of theta series of arbitrary Siegel degree as a finite double character sum over quotients of commensurable lattices. For the root lattice $A_2$, it yields a finite Fourier identity for cubes of cubic theta functions, a new proof of the two-variable cubic identity, and seven independent relations in degree two. A second specialization gives a new square analogue of the two-variable cubic identity. At the origin it reduces to $a(q)a(q^2)=b(q)b(q^2)+c(q)c(q^2)$; in degree two, the same seven relation patterns recur with each cube replaced by the corresponding product at $τ$ and $2τ$. We further extend the cubic identities on $A_2$ to $\ell$-th-power identities on $A_{\ell-1}$ for every $\ell\geq3$, and to $2\ell$-th-power identities on $A_{\ell-1}$ for every $\ell\geq2$, in arbitrary Siegel degree. The doubled-order construction recovers Jacobi's quartic identity and gives its analogues in arbitrary Siegel degree.

math.NT↗

arXiv2026

Efficient K-generalizable Learned Search

Learned top-K search improves the accuracy-latency trade-off of graph-based vector search, but existing methods target a fixed K. Serving real-world workloads with various K values thus incurs preprocessing cost proportional to the number of distinct Ks. We present OMEGA, which supports efficient arbitrary K learned search with the preprocessing cost of one top-1 model. It achieves so by reducing learned top-K search to repeated masked learned top-1 search with refinement, using carefully crafted trajectory features that generalize across Ks. OMEGA further employs rank-wise confidence allocation to tame error accumulation during refinements and applies a statistical forecast method to reduce model invocation overhead without compromising recall. Across nine dataset-scale configurations, OMEGA meets the 0.95 recall target and reduces mean latency by up to 65% (average: 23%; median: 20%) compared with learned search methods (including LAET, DARTH, MultiK-DARTH, ANNiE, and Ada-ef) at their lowest-preprocessing configurations. On production traces, it reduces total CPU time by up to 28% versus traditional HNSW search.

cs.DB↗

arXiv2026

Characterizing CPU-GPU Memory Sharing over Unified Memory Architecture

Unified Memory Architecture (UMA) enables the CPU and GPU to share a single physical memory. However, its complexity and opacity often drive applications back to inefficient copying or to incorrect direct access that sacrifices performance or even correctness. This paper presents a comprehensive study of CPU-GPU memory sharing on UMA, covering correctness and performance across data-plane transmission and control-plane coordination on four platforms. It yields four findings that guide libumsh, a portable library that delivers correct and efficient memory sharing across platforms and extends to emerging ones (e.g., NVIDIA DGX Spark). Applied to nine real-world systems, libumsh reduces data-plane time by 16-75% and control-plane overhead by 7-90%.

cs.OS↗

arXiv2026

Reduced-Rank Autoregressive Models for Matrix Time Series

Matrix time series is a series of matrix data observed over time. Analytical tools for such time series is needed in many applications in finance, economics, engineering and many other fields. To avoid the use of vectorization of the matrices which loses the column and row information, and the vector autoregression framework in traditional time series analysis, \cite{chen2021autoregressive} proposed the Matrix Autoregressive (MAR) Model. The model maintains and utilizes the matrix structure, leading to a substantial dimensional reduction and admitting explicit interpretations, comparing with the vector autoregressive model on the vectorized data. However, the MAR model still encounters difficulties in dealing with large dimensional matrix time series as the coefficient matrices in MAR models are also large. In this paper we propose to achieve further dimension reduction through reduced-rank constraints of the coefficient matrices in the MAR model. Estimation and rank determination procedures are studied. Theoretical investigation and empirical examples show that the reduced-rank constraint can achieve higher statistical efficiency than the MAR model.

stat.ME↗

arXiv2026

Proofs of two conjectures of Berkovich and Dhar on hook lengths

We prove stronger forms of two conjectures of Berkovich and Dhar concerning a coefficientwise inequality and pairs of hooks of length two in bounded partitions. The coefficientwise inequality implies the corresponding inequality for hook pairs. In particular, for each fixed size and common bound on the largest part, the total number of such pairs in odd partitions is at least that in distinct partitions. We give combinatorial proofs by explicit weight-preserving injections based on Glaisher's bijection and analytic proofs using finite product identities and known hook generating functions.

math.CO↗

arXiv2026

SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling

LLM scheduling is critical to serving, yet how well existing designs fit agentic serving--where agents, not humans, issue the requests--remains unclear. Agents shift the workload in two ways: they consume many more tokens than humans, so the cluster must provide high throughput (TPS) at low latency; and their requests reuse far more KV\$ than chat. Existing schedulers still trade off load balance against KV\$ reuse: cache-aware schedulers may crowd requests onto the few instances caching the KV\$, leaving the rest idle, while balanced schedulers may lose the opportunity for reuse, which is costly at a high reuse ratio. We thus present two key insights: (1) with a global-tier KV\$ store, pursuing load balance need not compromise KV\$ reuse, though the slower global tier must be used with care; and (2) given the agent's intra-session locality, routing requests by their sessions can balance the load with high KV\$ reuse. A key challenge in realizing session-centric scheduling is that the scheduler must identify a request's session statelessly, which is difficult for model providers serving arbitrary agents. SMetric addresses this with differential scheduling based on two indicators derived from the request itself, the session turn and the local KV\$ hit: it schedules first-turn requests for load balance, and sticks follow-ups to the instance with the highest local hit for high KV\$ reuse. As sessions differ widely in size, SMetric sticks a follow-up only if the instance can serve it within its SLO, and otherwise migrates the session to the least-loaded instance to prevent many long sessions from eventually imbalancing the load. Evaluated on real-world traces, SMetric improves the peak TPS by 9-15% under prefill-decode colocation with a provisioned global tier and the peak prefill TPS by 9% under disaggregation over state-of-the-art schedulers, also with lower latency.

cs.DC↗

arXiv2026

Tree-partitions of graphs with bounded tree-depth

Wood~ recently showed that every graph $G$ of pathwidth $h$ and $Δ(G)\ge1$ admits a $T$-partition of width at most $4(h+1)^2Δ(G)$ for some tree $T$ with $pw(T)\leq2h+1$. In this paper, we establish an analogous result for tree-depth, which is a stronger parameter than pathwidth. We prove that every connected graph with tree-depth $h$ admits a $T$-partition of width at most $\mathrm{max}\{1, (4h-10)Δ(G)+1\}$ for some tree $T$ with $\operatorname{rad}(T)\leq h-1$.

math.CO↗

arXiv2026

A Riemannian Factor Model for Manifold-Valued Time Series

We propose a Riemannian factor model (RFM), a novel framework for analyzing potentially high-dimensional time series data observed on Riemannian manifolds. Such time series are encountered in various applications, including economics, finance, medical imaging, and genomics and microbiome research. The proposed model is geometry-aware and accounts for the inherent nonlinearity in the data. In a high-dimensional asymptotic regime, where the manifold dimension is allowed to diverge with the sample size $n$, we establish convergence rates for the estimated loading space. In particular, under short-memory and strong factor conditions, we obtain a dimension-free $n^{-1/2}$ rate, which matches the convergence rate of the high-dimensional linear factor model. Finite-sample performance of the proposed RFM is demonstrated with simulated time series on the Bures--Wasserstein manifolds and products of spheres, as well as an application to monthly realized covariances of selected U.S. stock returns---modeled as time series in the Bures--Wasserstein manifold, where the RFM provides demonstrably interpretable factors and yields competitive predictive performance.

stat.ME↗

arXiv2026

Efficient and High-Accuracy Ray Tracing in Discretized Ionospheric Models

High-frequency (HF) ray tracing in complex ionospheric media generally involves a fundamental trade-off between path accuracy and computational efficiency, which directly affects practical applications such as over-the-horizon radar, ionospheric monitoring, and HF skywave communication systems. This paper presents RTM-GD, a ray-tracing framework that combines Hamiltonian ray integration with a continuously differentiable Galerkin--Difference (GD) interpolation strategy. Under discretized ionospheric grid conditions, the electron density is reconstructed as a $C^1$-continuous function within each grid cell, yielding an everywhere differentiable electron-density field for stable numerical integration and improved propagation-path accuracy. Numerical simulations and validations using measured HF oblique sounding data are conducted under diverse conditions, including different ionospheric states, low- and high-elevation angles, multiple operating frequencies, and both ordinary (O) and extraordinary (X) wave modes. Results show that RTM-GD consistently achieves sub-kilometer RMSE in both group-path and ground-distance metrics and sub-0.01-degree azimuth deviation relative to Richardson extrapolation, while reducing computational time by 98\%. Compared with Catmull--Rom interpolation, RTM-GD reduces the RMSEs of ray parameters by approximately one order of magnitude with less than 4\% additional computational cost. Measured-data validation based on ionogram synthesis further shows that the mean relative group-path error remains within 7\%, confirming reliable reproduction of practical HF oblique propagation characteristics. Overall, RTM-GD provides an accurate and computationally efficient framework for HF ray tracing in discretized ionospheric environments.

physics.ao-ph↗

arXiv2026

Transformation Formulae and Applications for Double Lambert Series

In this paper, we study a class of double Lambert series and establish several identities and transformation relations for them. These formulae provide useful tools for reducing certain double Lambert series to single Lambert series. As applications, we derive identities related to recent conjectures of Andrews, Dixit, Schultz, and Yee, and of Amdeberhan, Andrews, and Ballantine. We also propose a new proof of a result of Amdeberhan, Andrews, and Ballantine.

math.NT↗

arXiv2026

Two-colored generalized Frobenius partitions and minimal-excludant sums over bipartitions

Let $\cpsi_{2,a}(n)$ denote the number of $(2,a)$-colored Frobenius partitions of weight $n$, where the two rows have prescribed length difference. We study the two cases $a=0$ and $a=1$ and connect them with minimal-excludant statistics on bipartitions. Let $σ\mex_2(n)$ be the sum of the Lin--Liu bipartition minimal excludants over all bipartitions of $n$, and let $E_2(n)$ be the number of bipartitions whose two component minimal excludants are equal. For all $n\geq 0$, we give a combinatorial proof of \[ \cpsi_{2,0}(n)=2σ\mex_2(n) \qquad\text{and}\qquad \cpsi_{2,1}(n)=2σ\mex_2(n)-E_2(n). \] These identities give direct combinatorial interpretations of two-colored Frobenius partition functions in terms of bipartition minimal-excludant sums.

math.CO↗

arXiv2026

Simple is Better: Multiplication May Be All You Need for LLM Request Scheduling

High-quality LLM request scheduling requires meeting two key objectives: ensuring the routed instance has KVCache to accelerate request execution, and ensuring that the workload is balanced across instances. Achieving both objectives is challenging because pursuing one may compromise the other. Current approaches use various combinators (e.g., linear combinations) to compute a scheduling score that combines indicators for the two objectives. These approaches are complex: they either require significant workload-specific hyperparameter tuning or model-hardware-aware simulator development, yet could still lead to suboptimal performance. In this paper, we show that using a simple multiplication of two carefully chosen indicators: one KVCache-aware (new prefill tokens if routed to an instance) and one load-balancing-aware (current batch size of the instance), as the scheduling score (LMETRIC) can achieve both objectives simultaneously without any hyperparameter tuning. The key idea is that the simply multiplied score considers both objectives in a manner similar to a linear combination, but the original hyperparameters cancel out during comparison, so no tuning is needed to find the best parameters. The two indicators are chosen based on our analysis of LLM characteristics. Our extensive experiments show that this simple approach can reduce TTFT by 92% and 39%, and TPOT by 24% and 51%, compared to vLLM-v1 and an in-production scheduler on real-world workloads covering chatbots and coding agents. We also derive the mathematical conditions under which multiplication may fail, and find that such conditions are extremely rare in practice and can be detected (and mitigated) beforehand. LMETRIC has been deployed in production and canary release confirms its effectiveness

cs.DC↗

arXiv2026

A Generalization of the Amdeberhan-Andrews-Ballantine Conjecture

In this paper, we prove a generalization of a conjecture of Amdeberhan, Andrews, and Ballantine on double Lambert series. Motivated by a question raised by Cui, Kumar, and Singh concerning the existence of a generalization of this conjecture, we establish an identity in which the coefficients are given by the generalized divisor function $σ_k(n)$. As a special case, our result includes the original conjecture.

math.NT↗

arXiv2026

On tree decompositions whose trees are subgraphs

Fix $k \in \mathbb{N}$ and let $G$ be a connected graph with treewidth at most $k$. We say that $xy \notin E(G)$ is a {\em $k$-ghost-edge} of $G$ if for every tree decomposition $(T, \cB)$ of $G$ with width at most $k$, both $x$ and $y$ are contained in a bag of $(T, \cB)$. Moreover, if $G$ does not contain any $k$-ghost-edges, then $G$ is {\em $k$-ghost-free}. Hickingbotham proposed a conjecture that every connected $k$-ghost-free graph $G$ has a tree decomposition $(T, \cB)$ with width at most $k$ such that $T$ is a subgraph of $G$. In this paper, we prove that Hickingbotham's conjecture is false for all $k\geq3$.

math.CO↗