Searcharxiv⌕ Search

arXiv subjects

Fuxi Zhang

Publications and source records attributed to Fuxi Zhang.

14 recordsLinked to original sources
arXiv2026

Revisiting Salient Object Detection from an Observer-Centric Perspective

Salient object detection is inherently a subjective problem, as observers with different priors may perceive different objects as salient. However, existing methods predominantly formulate it as an objective prediction task with a single groundtruth segmentation map for each image, which renders the problem under-determined and fundamentally ill-posed. To address this issue, we propose Observer-Centric Salient Object Detection (OC-SOD), where salient regions are predicted by considering not only the visual cues but also the observer-specific factors such as their preferences or intents. As a result, this formulation captures the intrinsic ambiguity and diversity of human perception, enabling personalized and context-aware saliency prediction. By leveraging multi-modal large language models, we develop an efficient data annotation pipeline and construct the first OC-SOD dataset named OC-SODBench, comprising 33k training, validation and test images with 152k textual prompts and object pairs. Built upon this new dataset, we further design OC-SODAgent, an agentic baseline which performs OC-SOD via a human-like "Perceive-Reflect-Adjust" process. Extensive experiments on our proposed OC-SODBench have justified the effectiveness of our contribution. Through this observer-centric perspective, we aim to bridge the gap between human perception and computational modeling, offering a more realistic and flexible understanding of what makes an object truly "salient." Code and dataset are publicly available at: https://github.com/Dustzx/OC_SOD

cs.CV↗
arXiv2026

Think3D: Thinking with Space for Spatial Reasoning

While Vision-Language Models (VLMs) excel at 2D visual understanding, they remain constrained by 2D-centric paradigm that severely limits genuine 3D spatial reasoning. To bridge this gap, we introduce Think3D, a novel framework that equips VLM agents with interactive, 3D chain-of-thought reasoning capabilities. By integrating a suite of 3D manipulation tools, Think3D transforms perception into active spatial exploration, mirroring human geometric reasoning. Think3D consistently improves proprietary models, including GPT-4.1 and Gemini 2.5 Pro, across BLINK Multi-view, MindCube-1K, and VSI-Bench-Tiny. We further propose Think3D-RL to teach smaller open-weight models how to manipulate 3D space effectively. Using only final-answer rewards, without process supervision or handcrafted exploration trajectories, Think3D-RL enables Qwen3-VL-4B to autonomously learn effective 3D exploration strategies. After training, the model exhibits tool-use patterns similar to those of stronger proprietary models, while shifting the effect of 3D tool use on MindCube-1K from a performance drop to a substantial improvement. These results show that active exploration in 3D space provides an effective and general paradigm for improving spatial reasoning in multimodal agents. Code, models, and data are available at https://github.com/zhangzaibin/spagent.

cs.CV↗
arXiv2025

How Far are VLMs from Visual Spatial Intelligence? A Benchmark-Driven Perspective

Visual Spatial Reasoning (VSR) is a core human cognitive ability and a critical requirement for advancing embodied intelligence and autonomous systems. Despite recent progress in Vision-Language Models (VLMs), achieving human-level VSR remains highly challenging due to the complexity of representing and reasoning over three-dimensional space. In this paper, we present a systematic investigation of VSR in VLMs, encompassing a review of existing methodologies across input modalities, model architectures, training strategies, and reasoning mechanisms. Furthermore, we categorize spatial intelligence into three levels of capability, ie, basic perception, spatial understanding, spatial planning, and curate SIBench, a spatial intelligence benchmark encompassing nearly 20 open-source datasets across 23 task settings. Experiments with state-of-the-art VLMs reveal a pronounced gap between perception and reasoning, as models show competence in basic perceptual tasks but consistently underperform in understanding and planning tasks, particularly in numerical estimation, multi-view reasoning, temporal dynamics, and spatial imagination. These findings underscore the substantial challenges that remain in achieving spatial intelligence, while providing both a systematic roadmap and a comprehensive benchmark to drive future research in the field. The related resources of this study are accessible at https://sibench.github.io/Awesome-Visual-Spatial-Reasoning/.

cs.AI↗
arXiv2025

Cognitive-Aligned Document Selection for Retrieval-augmented Generation

Large language models (LLMs) inherently display hallucinations since the precision of generated texts cannot be guaranteed purely by the parametric knowledge they include. Although retrieval-augmented generation (RAG) systems enhance the accuracy and reliability of generative models by incorporating external documents, these retrieved documents often fail to adequately support the model's responses in practical applications. To address this issue, we propose GGatrieval (Fine-\textbf{G}rained \textbf{G}rounded \textbf{A}lignment Re\textbf{trieval} for verifiable generation), which leverages an LLM to dynamically update queries and filter high-quality, reliable retrieval documents. Specifically, we parse the user query into its syntactic components and perform fine-grained grounded alignment with the retrieved documents. For query components that cannot be individually aligned, we propose a dynamic semantic compensation mechanism that iteratively refines and rewrites the query while continuously updating the retrieval results. This iterative process continues until the retrieved documents sufficiently support the query's response. Our approach introduces a novel criterion for filtering retrieved documents, closely emulating human strategies for acquiring targeted information. This ensures that the retrieved content effectively supports and verifies the generated outputs. On the ALCE benchmark, our method significantly surpasses a wide range of baselines, achieving state-of-the-art performance.

cs.AI↗
arXiv2019

A large sample property in approximating the superposition of i.i.d. point processes

One of the main differences between the central limit theorem and the Poisson law of small numbers is that the former possesses the large sample property (LSP), i.e., the error of normal approximation to the sum of $n$ independent identically distributed (i.i.d.) random variables is a decreasing function of $n$. Since 1980's, considerable effort has been devoted to recovering the LSP for the law of small numbers in discrete random variable approximation. In this paper, we aim to establish the LSP for the superposition of i.i.d. point processes.

math.PR↗
arXiv2019

Liouville first passage percolation: geodesic length exponent is strictly larger than 1 at high temperatures

Let $\{η(v): v\in V_N\}$ be a discrete Gaussian free field in a two-dimensional box $V_N$ of side length $N$ with Dirichlet boundary conditions. We study the Liouville first passage percolation, i.e., the shortest path metric where each vertex is given a weight of $e^{γη(v)}$ for some $γ>0$. We show that for sufficiently small but fixed $γ>0$, with probability tending to $1$ as $N\to \infty$, all geodesics between vertices of macroscopic Euclidean distances simultaneously have (the conjecturally unique) length exponent strictly larger than 1.

math.PR↗
arXiv2018

Heat kernel for Liouville Brownian motion and Liouville graph distance

We show the existence of the scaling exponent $χ\in (0,4[(1+γ^2/4)- \sqrt{1+γ^4/16}]/γ^2]$ of the graph distance associated with subcritical two-dimensional Liouville quantum gravity of paramater $γ<2$ on $\mathbb V =[0,1]^2 $. We also show that the Liouville heat kernel satisfies, for any fixed $u,v\in \mathbb V^o$, the short time estimates $$ \lim_{ t \to 0} \frac{\log |\log {\mathsf p}_t^γ(u,v)| }{|\log t|}=\fracχ{2-χ}, \ \mbox{\rm a.s.} $$

math.PR↗
arXiv2017

Non-universality for first passage percolation on the exponential of log-correlated Gaussian fields

We consider first passage percolation (FPP) where the vertex weight is given by the exponential of two-dimensional log-correlated Gaussian fields. Our work is motivated by understanding the discrete analog for the random metric associated with \emph{Liouville quantum gravity} (LQG), which roughly corresponds to the exponential of a two-dimensional Gaussian free field (GFF). The particular focus of the present paper is an aspect of universality for such FPP among the family of log-correlated Gaussian fields. More precisely, we construct a family of log-correlated Gaussian fields, and show that the FPP distance between two typically sampled vertices (according to the LQG measure) is $N^{1+ O(ε)}$, where $N$ is the side length of the box and $ε$ can be made arbitrarily small if we tune a certain parameter in our construction. That is, the exponents can be arbitrarily close to $1$. Combined with a recent work of the first author and Goswami on an upper bound for this exponent when the underlying field is a GFF, our result implies that such exponent is \emph{not} universal among the family of log-correlated Gaussian fields.

math.PR↗
arXiv2017

On the Liouville heat kernel for k-coarse MBRW and nonuniversality

We study the Liouville heat kernel (in the $L^2$ phase) associated with a class of logarithmically correlated Gaussian fields on the two dimensional torus. We show that for each $\varepsilon>0$ there exists such a field, whose covariance is a bounded perturbation of that of the two dimensional Gaussian free field, and such that the associated Liouville heat kernel satisfies the short time estimates, $$ \exp \left( - t^{ - \frac 1 { 1 + \frac 1 2 γ^2 } - \varepsilon } \right) \le p_t^γ(x, y) \le \exp \left( - t^{- \frac 1 { 1 + \frac 1 2 γ^2 } + \varepsilon } \right) , $$ for $γ<1/2$. In particular, these are different from predictions, due to Watabiki, concerning the Liouville heat kernel for the two dimensional Gaussian free field.

math.PR↗
arXiv2011

Stein's method and locally dependent point process approximation

Random events in space and time often exhibit a locally dependent structure. When the events are very rare and dependent structure is not too complicated, various studies in the literature have shown that Poisson and compound Poisson processes can provide adequate approximations. However, the accuracy of approximations does not improve or may even deteriorate when the mean number of events increases. In this paper, we investigate an alternative family of approximating point processes and establish Stein's method for their approximations. We prove two theorems to accommodate respectively the positively and negatively related dependent structures. Three examples are given to illustrate that our approach can circumvent the technical difficulties encountered in compound Poisson process approximation [see Barbour & Månsson (2002)] and our approximation error bound decreases when the mean number of the random events increases, in contrast to increasing bounds for compound Poisson process approximation.

math.PR↗
arXiv2008

Polynomial birth-death distribution approximation in Wasserstein distance

The polynomial birth-death distribution (abbr. as PBD) on $\ci=\{0,1,2, >...\}$ or $\ci=\{0,1,2, ..., m\}$ for some finite $m$ introduced in Brown & Xia (2001) is the equilibrium distribution of the birth-death process with birth rates $\{α_i\}$ and death rates $\{β_i\}$, where $\a_i\ge0$ and $\b_i\ge0$ are polynomial functions of $i\in\ci$. The family includes Poisson, negative binomial, binomial and hypergeometric distributions. In this paper, we give probabilistic proofs of various Stein's factors for the PBD approximation with $\a_i=a$ and $\b_i=i+bi(i-1)$ in terms of the Wasserstein distance. The paper complements the work of Brown & Xia (2001) and generalizes the work of Barbour & Xia (2006) where Poisson approximation ($b=0$) in the Wasserstein distance is investigated. As an application, we establish an upper bound for the Wasserstein distance between the PBD and Poisson binomial distribution and show that the PBD approximation to the Poisson binomial distribution is much more precise than the approximation by the Poisson or shifted Poisson distributions.

math.PR↗
arXiv2005

On the speed of random walks on a percolation cluster of trees

We consider the simple random walk on the infinite cluster of the Bernoulli bond percolation of trees, and investigate the relation between the speed of the simple random walk and the retaining probability p by studying three classes of trees. A sufficient condition is established for Galton-Watson trees.

math.PR↗
arXiv2005

The reversible nearest particle systgems on a finite interval

In this paper we study a one-parameter family of attractive reversible nearest particle system on a finite interval. As the length of the interval increases, the time that the nearest particle system first hits the empty set increases in different order, from logarithmic to exponential, according to the intensity of interaction. In particular, at the critical case, the first hitting time increases in a polynomial order.

math.PR↗