arXiv subjects

Xiaolin Wang

Publications and source records attributed to Xiaolin Wang.

At least 19 recordsLinked to original sources

arXiv2026

RTP-LLM: Building an Evolvable and Reliable LLM Inference Engine at Alibaba Scale

Large language models and inference techniques evolve rapidly, making model enablement a recurring production task rather than a one-time integration effort. At Alibaba, a shared inference infrastructure must keep pace with new models and serving capabilities while remaining reliable across heterogeneous production environments. To reconcile rapid evolution with production reliability, we present RTP-LLM, an open-source production inference engine developed and deployed at Alibaba. Its architecture separates model-specific execution from shared request processing, state management, and resource coordination. Explicit interfaces localize compatible model changes while allowing models and deployment configurations to reuse common serving mechanisms. RTP-LLM has evolved from an exploratory system into a shared production engine that supports $\sim$100 models and processes hundreds of trillions of tokens per day. We describe how its architecture and production practices accommodate model evolution and contain the impact of failures, and share lessons for building evolvable and reliable inference engines at enterprise scale.

cs.OS↗

arXiv2026

Sharp Fourth-Moment Wasserstein CLT Bounds

We prove that for the normalized sum $W=\sum_{i=1}^n X_i/\sqrt{n}$ of independent, mean-zero, $d$-dimensional random vectors with $\operatorname{Cov}(W) =Σ=\operatorname{diag}(σ_1^2,\dots, σ_d^2)$ (the diagonal form of the covariance matrix is assumed without loss of generality), the 2-Wasserstein distance from $W$ to $N(0,Σ)$ is bounded by \[\sqrt{2\sum_{j=1}^d \min \left\{σ_j^2, \frac{\sum_{i=1}^n \mathbb{E} [|X_i'-X_i|^2(X'_{ij}-X_{ij})^2 ] }{n^2 σ_j^2} \right\}},\] where $|\cdot|$ denotes the Euclidean norm, $X_i'$ is an independent copy of $X_i$, and $X_{ij}$ ($X_{ij}'$ resp.) denotes the $j$th coordinate of the vector $X_i$ ($X_i'$ resp.). On the one hand, it gives the optimal rate $O(β\sqrt{d/n})$ if $|X_i|\leq β$ for all $i$. On the other hand, in the isotropic case where $Σ=I_d$, it gives the optimal fourth-moment bound $O\left(\sqrt{\sum_{i=1}^n \mathbb{E}|X_i|^4/n^2}\right)$. Moreover, the coordinatewise minima suggest a tradeoff between the naïve coupling and the normal coupling for each component. To prove the main result, we develop a new transportation inequality using a coupling via the Föllmer process with a general covariance matrix and the martingale property of its drift. The resulting bound is then controlled using a Bonis-type expansion together with Bessel's inequality. The method and the result extend to $p$-Wasserstein bounds. As a consequence, we obtain the Hilbert space 4-Wasserstein CLT bound $O(n^{-1/4}[\operatorname{tr}(Q^{1/2})]^{1/2})$, where, in the i.i.d. case, $Q=\mathbb{E}[|X_1'-X_1|^2(X_1'-X_1)^{\otimes2}]$ and the $O(n^{-1/4})$ rate is optimal given the dimension-free form of $Q$. We further strengthen it to a dimension-free strong approximation result with a near-optimal error rate through the KMT embedding construction.

math.PR↗

arXiv2026

Induced Size-Ramsey Numbers of Odd Cycles via Spread Measures

We prove that, for every $k\ge2$ and every sufficiently large odd $n$, the induced size--Ramsey number of $C_n$ satisfies \[ \widehat R_{\mathrm{ind}}(C_n;k)=e^{Θ(k)}n. \] This improves the previous upper bound of $e^{O(k\log k)}n$ and resolves a conjecture of Bradač, Draganić and Sudakov. We obtain this result by proving a stronger form of their odd-hole gadget conjecture: we construct a graph with $e^{O(k)}$ edges whose every $k$-edge-colouring contains a monochromatic induced odd cycle of length $O(\log k)$. Our proof develops a spread-based approach to finding induced cycles. The key idea is to place spread probability weights directly on families of candidate cycles and control their possible chords through a quadratic chord score. Combining this with random sparsification and hypergraph containers yields a sparse host graph that works for every $k$-edge-colouring.

math.CO↗

arXiv2026

High-dimensional normal approximations for sums of Langevin Markov chains

Consider the well-known Langevin diffusion on $\mathbb{R}^d$ $$\mathrm{d} X_t = -\nabla U(X_t)\,\mathrm{d} t + \sqrt{2}\mathrm{d} B_t, $$ and its Euler-Maruyama discretization given by $$X_{k+1}=X_k-η\nabla U(X_k)+\sqrt{2η}ξ_{k+1},$$ where $η$ is the step size. Under mild conditions, the Langevin diffusion admits $π(\mathrm{d} x)\propto \exp(-U(x))\mathrm{d} x$ as its unique stationary distribution. In this paper, we mainly study the normal approximation of the normalized partial sum $$ W_n = η^{1/2} n^{-1/2} \left( \sum_{i=0}^{n-1} X_i- \int_{\mathbb{R}^d} x\,π(\mathrm{d} x) \right).$$ To the best of our knowledge, this work provides the first dimension-explicit convergence rates in high-dimensional settings. Our main tool is a novel upper bound for the 1-Wasserstein distance $W_1(W,γ)$ via the exchange pair approach, where $W$ is any random vector of interest and $γ$ is a $d$-dimensional standard normal random vector.

math.PR↗

arXiv2026

Jianmu-Q: Bridging Quantum Field Theory and Quantum Hardware with AI Agents

Quantum computing offers a promising approach to non-perturbative simulations of quantum field theories, particularly for real-time dynamics and systems affected by the sign problem. Yet translating continuum field theories into quantum computations remains a challenging interdisciplinary task. Here we introduce \texttt{Jianmu-Q}, a multi-agent framework that transforms natural-language descriptions of quantum field theories into quantum simulation workflows. Specialized agents handle theoretical formulation, lattice discretization, qubit mapping, and circuit design, while physical consistency checks and independent numerical calculations provide validation throughout the workflow. We benchmark the framework on three field theories of increasing complexity: the $(1+1)$-dimensional Nambu--Jona-Lasinio and Schwinger models, and a $(2+1)$-dimensional $\mathbb{Z}_2$ gauge theory. Across these systems, noiseless quantum-circuit simulations converge toward independent exact-diagonalization and tensor-network results, while selected circuits are executed on IBM quantum hardware, demonstrating an end-to-end, validation-guided approach to quantum simulation of field theories.

hep-lat↗

arXiv2026

A quantitative CLT for the zeros of random trigonometric polynomials

Let $X_N(t)$, $t\in[0,π]$, be a stationary Gaussian trigonometric polynomial, and let $\mathcal Z_N$ denote the number of zeros of the rescaled process $X_N(t/N)$ on $[0,Nπ]$. We establish the quantitative central limit theorem \[ d_{\mathrm W}\left( \frac{\mathcal Z_N-\mathbb E[\mathcal Z_N]} {\sqrt{\operatorname{Var}(\mathcal Z_N)}},Z\right) \le C\frac{(\log N)^2}{\sqrt N}, \] where $Z$ is standard Gaussian and $d_{\mathrm W}$ denotes the Wasserstein-1 distance. The proof combines Ornstein--Uhlenbeck regularization of the singular Kac representation with direct estimates of the covariance between the Malliavin--Stein factor and Stein test functions. A lattice lower bound of order $N^{-1/2}$ shows that the convergence rate is optimal up to logarithmic factors.

math.PR↗

arXiv2026

Sharp linear Turán remainders for cycle and edge extremality

For a graph $G$, let $e(G)$ denote the number of edges of $G$, and let $c(G)$ be the number of distinct cycles in $G$. Morrison, Roberts and Scott asked whether, for every fixed graph $H$ and all large $n$, some $n$-vertex $H$-free graph maximizes both $e(G)$ and $c(G)$. In this paper, we show that the answer is no for every possible chromatic number. Let $T_{n,r}$ be the complete $r$-partite Turán graph on $n$ vertices. We find a constant $γ_r>0$ such that if $n$ is sufficiently large and $\mathcal{H}$ is a finite family with $χ(\mathcal{H})=r+1$ satisfying \[ ex(n,\mathcal{H})<e(T_{n,r})+γ_r n, \] then every cycle-maximal $\mathcal{H}$-free graph is edge-extremal, and $γ_r$ is best possible.

math.CO↗

arXiv2026

An asymptotically tight bound on oriented diameter

Let $f(d)$ be the smallest integer such that every finite connected bridgeless simple graph of diameter $d$ admits a strong orientation of diameter at most $f(d)$. Chvátal and Thomassen (1978) proved $\lceil d^2/2\rceil+d\le f(d)\le2d^2+2d$. In this paper, we prove that $\lceil d^2/2\rceil+d\le f(d)\le\lceil d^2/2\rceil+d+18$ for every $d\ge2$, which shows that $f(d)=\lceil d^2/2\rceil+d+O(1)$ and determines both the quadratic and linear terms up to a bounded additive error. The key method of our proof is to construct a central subgraph $H$ that admits a strong orientation of diameter $O(d)$ and is within distance $\lfloor d/2\rfloor$ of every vertex outside it. We obtain the sharp linear coefficient by jointly estimating outside paths and their connecting paths in $H$ at the actual attachment vertices, rebuilding $H$ when necessary.

math.CO↗

arXiv2026

An asymptotically tight bound on oriented diameter

Let $f(d)$ be the smallest integer such that every connected bridgeless graph of diameter $d$ admits an orientation with directed diameter at most $f(d)$. Chvátal and Thomassen (1978) proved $\left\lceil{d^2}/{2}\right\rceil+d\le f(d) \le 2d^2+2d$. In this paper, we prove that $f(d)\le \left\lceil d^2/2\right\rceil+6d$, which shows $f(d)= d^2/2+Θ(d)$ and determine the leading quadratic coefficient of $f(d)$. The key method of our proof is to construct a central subgraph $H$ such that it admits a strong orientation of diameter $O(d)$ and the distance between $H$ and every vertex outside $H$ is at most $ d/2 $.

math.CO↗

arXiv2026

Denoising-Enhanced Coarse-to-Fine Infrared Small Target Detection with Attention Prior-Guided Knowledge Distillation

Infrared small target detection (IRSTD) in high-resolution images is crucial for unmanned aerial vehicle (UAV) surveillance and UAV-based ground monitoring. However, small target size, weak features, and interference from complex dynamic backgrounds make IRSTD challenging. Existing methods incur redundant computation in non-target background regions and insufficiently exploit target context, limiting detection performance. To address these issues, we propose ECFNet, an efficient coarse-to-fine IRSTD framework with attention prior-guided knowledge distillation. In the coarse stage, we design a region binary classification network (RBCN) on grid-based multi-scale feature maps to efficiently identify target-containing context region proposals. A new denoising-assisted training strategy incorporates noisy ground-truth (GT) masks into RBCN feature maps and trains the network to reconstruct the original GT masks. This auxiliary task encourages explicit learning of target-background context to better distinguish target proposals from background regions. In the fine stage, we customize a lightweight target detector to the coarse-stage region proposals to balance accuracy and efficiency. Furthermore, we introduce a knowledge distillation strategy guided by a teacher-student cross-attention prior. This strategy directs the student to focus on critical target regions, enhancing discriminative feature representations for infrared small targets. Extensive experiments on three real infrared datasets demonstrate that ECFNet outperforms existing single-stage and two-stage approaches while maintaining high real-time processing efficiency. Code: https://github.com/IVPLabs/ECFNet.

cs.CV↗

arXiv2026

Semileptonic $\bar{B}^*_{c} \to (P, V) \ell^- \barν_{\ell}$ decays within the self-consistent light-front quark model

Motivated by the recent observation of the $B_c^*$ meson by the ATLAS Collaboration and its promising discovery potential at future high-luminosity colliders, we investigate the semileptonic decays $\bar{B}^*_{c} \to (P, V) \ell^- \barν_{\ell}$ ($P=D$, $B_{d,s}$, $η_c(1S,\,2S,\,3S)$; $V=D^*$, $B^*_{d,s}$, $ψ(1S,\,2S,\,3S)$) in the Standard Model. Using the relevant form factors obtained in the self-consistent covariant light-front quark model, we present theoretical predictions for several physical observables, including the branching fraction, the ratios of branching fractions $R_{(L)}$, the longitudinal polarization fraction of the daughter vector meson, the $τ$ lepton spin asymmetry and the forward-backward asymmetry. It is found that the branching fractions for the Cabibbo-favored $\bar{B}^*_{c}$ $\to$ $\bar{B}^*_{s} e^- \barν_e$ and $\bar{B}^*_{s} μ^- \barν_μ$ decays reach the order of ${\cal O}(10^{-6})$, which might be firstly observed at the future high-luminosity Large Hadron Collider.

hep-ph↗

arXiv2026

Mixing Time of Conditional Two Star Exponential Random Graphs

Two star exponential random graph models (ERGMs) are an interesting special case of both general ERGMs and mean-field Ising models. In this paper, we study two star ERGMs conditioning on the edge density $p$. We begin with an analytic characterization of the replica symmetric region, where the conditional model is close in cut distance to the Erdős--Rényi random graph $G(n,p)$. We prove that this region is twice as large as the corresponding region for the unconditional model. To study refined properties of the conditional model, we then analyze the global Kawasaki algorithm for sampling from it. Within the replica symmetric region, and under the additional condition that $4βp(1-p)<0.5$ when $|p-1/2|\lesssim 0.4632$, where $β$ is the model parameter, we prove metastable fast mixing of the Kawasaki algorithm. As corollaries, we obtain a weak Poincaré inequality and use it to deduce concentration inequalities of optimal order for conditional subgraph counts. The additional condition $4βp(1-p)<1/2$ if $|p-1/2|\lesssim 0.4632$ comes from our proof technique of contractive coupling. This bottleneck did not appear in previous studies applying the contractive coupling technique to unconditional ERGMs.

math.PR↗

arXiv2026

Adaptive Context Parallelism for Production LLM Serving

As LLM context windows expand and input sequences grow longer, serving systems face increasing computational and memory demands. Context parallelism (CP), which partitions the input sequence across multiple ranks to parallelize the computation, has therefore become increasingly important for efficient LLM serving. However, existing CP-enabled systems either rely on static CP configurations or adjust the CP degree only for active requests or batches. In this paper, we present Vertumnus, an adaptive CP serving system designed for heterogeneous and evolving workloads. At the request level, Vertumnus routes requests among workers with different CP degrees using a placement cost that combines predicted queuing delay, cache-aware prefill time, and GPU-time cost. At the cluster level, Vertumnus adapts the worker composition through seconds-scale split and merge operations as workload demand changes. Vertumnus further introduces a global prefix-cache management policy that coordinates cache placement and replication among workers with the same or different CP degrees, preserving cache locality as request assignments and worker composition change. Experiments on a 64-GPU cluster with public and production workloads show that, under the highest evaluated loads, Vertumnus reduces mean TTFT by up to 28.1% and improves token-weighted SLO attainment by up to 13.3 percentage points over the strongest baseline.

cs.OS↗

arXiv2026

The Multicolour Size--Ramsey Number of an Even Cycle

We determine the $k$-colour size--Ramsey number of even cycles up to absolute constant factors. For every $k\ge2$ and every even $n\ge100\log k$, \[ \widehat R_k(C_n)=Θ(k^2\log k)n. \] The lower bound follows from the corresponding result of Beke, Li and Sahasrabudhe for paths, while our upper bound improves the previous best estimate $O(k^{34}n)$ of Javadi, Kohayakawa and Miralaei.

math.CO↗

arXiv2026

Clique-saturating non-edges throughout the Turán range

For an $F$-free graph $G$, a non-edge is $F$-saturating if adding it to $G$ creates a copy of $F$. We denote by $f_{p+1}(n,m)$ the minimum number of $K_{p+1}$-saturating non-edges in a $K_{p+1}$-free $n$-vertex graph with $m$ edges. Erdős and Tuza conjectured that $f_4\left(n,\mathrm{ex}(n,K_3)+ 1\right)= (1 + o(1)) \frac{n^2}{16}$. Balogh and Liu (JCTB, 2014) disproved this conjecture and determined the asymptotic value of $f_4(n,\mathrm{ex}(n,K_3)+1)$. He, Ma, Ma and Ye (JCTB, 2023) later determined $f_{p+1}(n,\mathrm{ex}(n,K_p)+1)$ asymptotically for every $p\ge 3$, and asked for the value of $f_{p+1}(n,m)$ for all $\mathrm{ex}(n,K_p)+1\le m\le \mathrm{ex}(n,K_{p+1})$ and every $p\ge 3$. In this paper, we answer their question asymptotically for all $\mathrm{ex}(n,K_p)+1\le m\le \mathrm{ex}(n,K_{p+1})$ and every $p\ge 3$. We also determine the exact value of $f_3(n,m)$ for all $0\le m\le \mathrm{ex}(n,K_3)$ by a different method.

math.CO↗

arXiv2026

Odd and even cycle lengths, minimum degree and chromatic number in graphs

We present the relations between clique number and chromatic number with given the number of odd or even or all cycle lengths. Let $L_o(G)$ be the set of odd cycle lengths of $G$ and $\ell_o(G)$ be the longest odd cycle length. Gyárfás (DM, 1992) proved a classic result: $χ(G)\leq 2|L_o(G)|+2$, and if $w(G)\leq 2|L_o(G)|+1$, then $χ(G)\leq 2|L_o(G)|+1$. Later, Wang (SIAM DM, 2008), Ma and Ning (SIAM DM, 2018) together determined the exact chromatic number when $|L_o(G)|=2$. We further prove that if $w(G)\leq 2|L_o(G)|$ for any $|L_o(G)|\geq 2$, then $χ(G)\leq 2|L_o(G)|$. We also construct a class of graphs with $w(G)=2|L_o(G)|-1$ but $χ(G)=2|L_o(G)|$ for every $|L_o(G)|\geq 2$. Using our result, we give a short proof of the relation between $w(G)$ and $χ(G)$ with given $\ell_o(G)$ proved by Kenkre and Vishwanathan (JGT, 2006). Let $L_e(G)$ be the set of even cycle lengths of $G$ and $\ell_e(G)$ be the longest even cycle length. Mihók and Schiermeyer (DM, 2004) proved that $χ(G)\leq 2|L_e(G)|+3$, and if $w(G)\leq 2|L_e(G)|+2$, then $χ(G)\leq 2|L_e(G)|+2$. We further prove that if $w(G)\leq 2|L_e(G)|+1$ and $|L_e(G)|\geq 3$, then $χ(G)\leq 2|L_e(G)|+1$. We also construct a class of graphs with $w(G)=2|L_e(G)|$ but $χ(G)=2|L_e(G)|+1$ for every $|L_e(G)|\geq 2$. Our result can deduce the relation between $w(G)$ and $χ(G)$ with given $\ell_e(G)$. Combining all the above results, we deduce the similar relations between $w(G)$ and $χ(G)$ with given the number of cycle lengths or the longest cycle length.

math.CO↗

arXiv2026

Two-block cycles and chromatic number of Hamiltonian digraphs

Let $k$ and $\ell$ be positive integers. The family $C(k,\ell)$ consists of all digraphs obtained from two internally vertex-disjoint directed paths of lengths at least $k$ and $\ell$, respectively, and identifying their initial vertices and their terminal vertices. Addario-Berry, Havet and Thomassé (JCT-B, 2007) asked whether, for any positive integers $k$ and $\ell$ with $k+\ell \ge 4$, the chromatic number $χ(D)$ is at most $k+\ell-1$ for every $C(k,\ell)$-free strongly connected digraph $D$. Let $D$ be a $C(k,\ell)$-free Hamiltonian digraph. Kim, Kim, Ma and Park (JGT, 2018) showed that $χ(D) \le k+\ell$ and the bound is attained when $k+\ell=5$. In this paper, we prove that $χ(D) \le k+\ell-1$ for $k+\ell\ge 6$ and this bound is best possible for all $k+\ell\geq 6$, which resolves the problem posed by Addario-Berry, Havet and Thomassé for Hamiltonian digraphs.

math.CO↗

arXiv2026

The maximum number of cycles of a given length in a nonhamiltonian graph

In 2026, Li and Zhan characterized the nonhamiltonian graphs of order $n$ with the maximum number of paths of length $k$, where $n$ and $k$ are integers satisfying $1\leq k\leq n-1$. This work solves and generalizes a problem proposed by Erdős in 1980. In this paper, we further determine the nonhamiltonian graphs of order $n$ attaining the maximum number of cycles of length $k$ for given integers $n$ and $k$ with $3\leq k\leq n-1$. As a corollary, we determine the generalized Turán number $\mathrm{ex}(n,C_k,C_n)$ for every $3\leq k\leq n-1$.

math.CO↗