Searcharxiv⌕ Search

arXiv subjects

Yi Han

Publications and source records attributed to Yi Han.

At least 19 recordsLinked to original sources
arXiv2026

Poisson statistics for one-dimensional Anderson--Bernoulli Schrödinger operators

For a one-dimensional discrete Schrödinger operator with bounded nonconstant iid potential, we prove a unit-intensity Poisson limit for local eigenvalues unfolded by the integrated density of states. The limit holds at every energy in the interior of the almost-sure spectrum. In particular, it holds for Bernoulli potentials at every fixed nonzero coupling. The proof uses auxiliary phase functions built from inverse stationary measures to obtain an exact IDS drift. Projective contraction then yields exponential crossing times, and a spatial subdivision argument gives independent Poisson counts in disjoint energy windows. At energies where the IDS has a finite positive derivative, the usual linear rescaling gives the same limit.

math-ph↗
arXiv2026

ANGLER: Verifier-Grounded Benchmarking with an Adaptive Attacker for Web Agents

Web agents are increasingly deployed to complete tasks autonomously on real websites, where adversarially injected content can induce them to perform unauthorized actions. Existing evaluations primarily use fixed attacks or focus on a single attack type, and evaluate attacks at the text level rather than through real state changes, thus underestimating how vulnerable agents are to diverse attacks. To address the gap, we introduce REEF, a benchmark of red-teaming executable environments with faithful verification for measuring the security vulnerabilities of web agents. REEF covers 9 verifier-grounded websites spanning travel, commerce, productivity, media, and communication, and pairs every authentic benign task with a malicious objective delivered through 5 attack types at both the user and the environment level. REEF's state-based verifiers provide the execution feedback that drives ANGLER, an adaptive near-miss-guided lure evolution framework for red-teaming. Rather than evaluating a fixed attack set, ANGLER uses outcomes from prior rollouts to guide LLM-based attack mutation and candidate selection, prioritizing near-miss seeds during refinement. Across 8 VLM-based web agents, ANGLER achieves an end-to-end attack success rate (ASR) of 20.36%, compared with 15.24% for the non-adaptive setting. Compared with attack-free execution on the same benign tasks, attacks generated by ANGLER reduce benign task success rate (BSR) from 61.9% to 44.4%.

cs.CR↗
arXiv2026

HIPC: Humanoid Iterative Planning and Control via Goal-Conditioned Motion Closure

Recent generative models are widely used to provide whole-body motion for humanoid tracking. Current methods combine motion generators with controllers in an open-loop paradigm, which leads to critical issues: tracking errors can accumulate into spatial drift, and the planner cannot adapt to unexpected physical disturbances without feedback from physical execution. However, directly feeding realized states back to the planner in a closed loop can cause discontinuities, resulting in physical jitter and even falls, since such states are out of distribution during training. To bridge this gap, we propose HIPC, a real-time Humanoid Iterative Planning and Control framework that couples a generative model with motion tracking into a unified planning-and-control paradigm. HIPC introduces a goal-conditioned motion planning trained with a canonical motion representation that avoids gimbal-lock singularities under extreme orientations and supports motion generation on diverse geometric and semantic conditions. Then, HIPC adopts two strategies, deviation-decayed recovery and history re-anchoring, to mitigate the out-of-distribution shift caused by realized-state feedback. Finally, HIPC operates a real-time asynchronous planning-and-control system, with latency compensation to reduce temporal mismatch for inference delay. Extensive experiments confirm the effectiveness of HIPC under switching and composition of diverse conditions. Results show that HIPC achieves state-of-the-art goal-reaching accuracy among existing methods while simultaneously demonstrating robustness to physical perturbations and the ability to recover from falls.

cs.RO↗
arXiv2026

Circular law for non-Hermitian block band matrices with slowly growing bandwidth

We consider the empirical eigenvalue distribution for a class of non-Hermitian random block tridiagonal matrices $T$ with independent entries. The matrix has $n$ blocks on the diagonal and each block has size $\ell_n$, so the whole matrix has size $n\ell_n$. We assume that the nonzero entries are i.i.d. with mean 0, variance $(3\ell_n)^{-1}$ and having finite moments of all orders. We prove that when the entries have a bounded density, then whenever $\lim_{n\to\infty}\ell_n=\infty$ and $\ell_n=O(\operatorname{Poly}(n))$, the normalized empirical spectral distribution of $T$ converges almost surely to the circular law. The growing bandwidth condition $\lim _{n\to\infty}\ell_n=\infty$ is the optimal condition for the circular law with small bandwidth. This confirms the folklore conjecture that the circular law holds whenever the bandwidth increases with the dimension; earlier circular-law proofs required the delocalized regime $\ell_n\gg n$.

math.PR↗
arXiv2026

The smallest singular value for rectangular random matrices with Lévy entries

Let $X=(x_{ij})\in\mathbb{R}^{N\times n}$ be a rectangular random matrix with i.i.d. entries (we assume $N/n\to\mathbf{a}>1$), and denote by $σ_{min}(X)$ its smallest singular value. When entries have mean zero and unit second moment, the celebrated work of Bai-Yin and Tikhomirov show that $n^{-\frac{1}{2}}σ_{min}(X)$ converges almost surely to $\sqrt{\mathbf{a}}-1.$ However, little is known when the second moment is infinite. In this work we consider symmetric entry distributions satisfying $\mathbb{P}(|x_{ij}|>t)\asymp t^{-α}$ for some $α\in(0,2)$, and prove that $σ_{min}(X)$ can be determined up to a log factor with high probability: for any $D>0$, with probability at least $1-n^{-D}$ we have $$C_1n^{\frac{1}α}(\log n)^\frac{2(α-2)}α\leq σ_{min}(X)\leq C_2n^{\frac{1}α}(\log n)^\frac{α-2}{2α}$$ for some constants $C_1,C_2>0$. The upper bound was derived in a recent work of Bao, Lee and Xu \cite{bao2024phase2} but the lower bound is new and answers a problem posed in that paper in a weaker form. This appears to be the first determination of $σ_{min}(X)$ in the $α$-stable case with a correct leading order of $n$, as previous anti-concentration arguments only yield lower bound $n^\frac{1}{2}$. The same lower bound holds for $σ_{min}(X+B)$ for any fixed rectangular matrix $B$ with no assumption on its operator norm. The case of diverging aspect ratio is also computed.

math.PR↗
arXiv2026

Recognition Is Not Grounding: Benchmarking Multilingual Financial Evidence Extraction

Comparing financial disclosures across languages requires identifying shared accounting concepts and recovering all numerical facts that express them. Existing benchmarks evaluate aspects of these capabilities, but provide limited insight into whether both succeed on the same multilingual evidence. In particular, it remains unclear whether language models can recognize these concepts across languages and exhaustively recover their corresponding numerical evidence from the same contexts. We introduce MULTIFINGROUND, a benchmark designed to investigate this recognition-grounding gap in multilingual financial reports. Constructed from FY2025 disclosures of 90 issuers in English, French, and Greek, MULTIFINGROUND aligns independently authored reports through a shared accounting taxonomy and provides traceable fact-level annotations. The benchmark evaluates financial fact identification, concept recognition, exhaustive fact selection, and concept discrimination without candidate labels. Across 15 evaluated models, eight API and open-weight reference models achieve 97.3-99.4% concept recognition accuracy, yet miss at least one target fact in 26.6-41.9% of correctly recognized groups. Further analyses reveal substantial variation in concept discrimination and limited gains from the evaluated finance-specialized models. Our findings demonstrate that strong concept recognition can coexist with incomplete evidence recovery, highlighting the importance of evaluating semantic recognition and evidence completeness together.

cs.AI↗
arXiv2026

Herculean: An Agentic Benchmark for Financial Intelligence

As AI agents advance, the central question is no longer whether they can solve isolated financial tasks, but whether their capabilities generalize across complete professional workflows. Existing benchmarks cannot answer this question because they evaluate either static competencies or agents within a single interaction structure. We introduce Herculean, a benchmark of end-to-end agentic financial intelligence that crosses agent architecture, backbone model, and four structurally distinct workflows: Trading, Hedging, Market Insights, and Auditing. Each workflow is implemented as an MCP skill that preserves its native state, actions, constraints, and outcome criteria, allowing intact agent systems to be evaluated under matched external conditions. Across 30 architecture--backbone configurations, we find that agent capability is not globally ordered. Workflow rankings are weakly aligned, changing the architecture frequently reverses performance under the same backbone, and each architecture concentrates failure in different forms of workflow control. These results establish that agent performance is an interaction between architecture and workflow demand, rather than an intrinsic property of the underlying model. Consequently, success on a single financial workflow provides little evidence of dependable, general financial intelligence.

cs.AI↗
arXiv2026

Degrees of Freedom in Multi-Agent Coordination: Between Centralized and Decentralized Regimes

Large language model (LLM)-based multi-agent systems (MAS) are commonly organized around two coordination regimes. Centralized coordination provides explicit control but can create a bottleneck at a single coordination locus, whereas decentralized coordination offers local autonomy but must achieve effective collaboration through distributed decisions and experience. These trade-offs motivate a systematic study of the space between the two regimes. We introduce Degrees of Freedom (DoF), a continuous configuration framework that quantifies the size of the decision space delegated to downstream agents. DoF controls the granularity of delegated work and the target probability that each task unit is assigned to the branches. Its endpoints recover the two coordination regimes. Across multiple task domains, DoF reveals a hierarchy of organizational effects. Delegated granularity and within-regime routing produce structured but bounded profile reallocations, whereas task domain and retained experience are associated with stronger differences between the centralized- and decentralized-branch profiles. The resulting performance landscape is task-dependent: different domains favor different DoF configurations, while atomic-target conditions are more consistent across runs. In-domain retained experience brings the two branch profiles closer together. The same controlled sweep reveals a task-dependent coordination landscape: some domains exhibit higher-performing configurations, whereas others show flatter responses and no single configuration dominates across domains. In five of the six evaluated state, an intermediate DoF configuration achieves higher mean accuracy than both endpoints, demonstrating the value of exploring the coordination space between them. Together, these findings provide empirical guidance for studying and designing intermediate coordination regimes in LLM-based MAS.

cs.MA↗
arXiv2026

The circular law for random band matrices with arbitrary doubly stochastic variance profiles

We study the circular law for non-Hermitian random band matrices with independent entries and arbitrary doubly stochastic variance profiles. For matrices of size $n$ and bandwidth $n^γ$, we prove the circular law when $γ>\frac{5}{6}$ for Gaussian entries, or $γ>\frac{8}{9}$ for subgaussian entries under bounded-density assumptions. These thresholds improve the earlier sufficient conditions $γ>\frac{32}{33}$ for block band matrices and $γ>\frac{33}{34}$ for periodic band matrices. Subsequent work proves the circular law under the optimal condition of diverging bandwidth for specific open-boundary models \cite{han2025circular2} and periodic models \cite{han2026optimal}. Those results rely on local band structure and do not cover the arbitrary doubly stochastic variance profiles considered here. We also derive weak delocalization estimates for right eigenvectors in the Gaussian case. Our main technical input consists of polynomial lower bounds for certain intermediate small singular values, requiring no structural assumptions on the variance profile beyond double stochasticity. We also prove a circular law for the root empirical spectral measure of products of an arbitrary number of i.i.d. matrices, under bounded-density assumptions and a finite fourth moment. For smaller numbers of factors, we apply the same strategy as in our main band-matrix theorem to the cyclic linearization. For larger numbers of factors, the simpler structure of the product model allows us to apply the singular-value universality theorem of Hanin and Jiang \cite{hanin2025global} directly, together with a shifted log-determinant comparison. The two methods apply in overlapping ranges and together cover all growth rates of the number of factors.

math.PR↗
arXiv2026

EviScene: Beyond "Is It AI?" for Agentic Image Factuality Assessment

Determining whether an image is AI-generated does not establish whether the situation it depicts is factually supported. We formulate Image Factuality Assessment: given an image and a generic verification instruction, a system returns a supported or refuted verdict with an evidence-grounded explanation. Without a supplied textual claim, the system must progressively refine factual questions during investigation and align external evidence with the specific situation depicted. We introduce EviLens, an expert-reviewed benchmark of 10,174 supported and refuted images from controlled construction and web collection, covering twelve factual error types and evaluating both verdict accuracy and evidence sufficiency. We develop EviScene, a multimodal ReAct agent that interleaves visual inspection, external retrieval, and evidence analysis, using new findings to refine its questions and revisit relevant image details. We train a compact policy with supervised fine-tuning on reviewed teacher trajectories, followed by privileged on-policy self-distillation applied to verified decisions from the policy's own investigations. On the 1,527-image test set, the trained Qwen3.5-9B agent achieves 79.05% balanced accuracy and a Strict Evidence Sufficiency Rate (SESR) of 52.26%, improving over the same-backbone base agent by 6.71 and 9.37 percentage points, respectively. Its SESR is the highest among the evaluated agents.Code, models, and data will be released at https://github.com/wangzian828/EVISCENE and https://huggingface.co/ZiAnwang/EviScene.

cs.CV↗
arXiv2026

Large deviations of the spectral radius of iid subgaussian random matrices

We study large deviations of the spectral radius of $X_n=n^{-1/2}(ξ_{ij})_{i,j=1}^n$, where the entries are iid, centered, variance-one, and subgaussian, with zero pseudovariance in the complex case. We write $ρ(X_n)$ for its spectral radius, i.e. the largest modulus among its eigenvalues. For every fixed $r>1$, we prove $$ \liminf_{n\to\infty}\frac1n\log\mathbb{P}\{ρ(X_n)>r\} \ge -\frac\beta2\bigl(r^2-1-2\log r\bigr), $$ where $β=1$ for real entries and $β=2$ for complex entries. We prove a matching upper bound for real symmetric laws with Gaussian-dominated even moments and for complex laws satisfying the sharp Gaussian Laplace-transform bound. These classes include discrete distributions. In the complex sharp class, we also identify the exponential rate for an eigenvalue to enter a fixed disk outside the unit disk. We then turn to lower deviations and establish a quadratic-speed bound. Under the additional assumption that the entry law $μ$ has a bounded density, we prove $$ \mathbb{P}\{ρ(X_n)\le r\}\le e^{-c_{μ,r}n^2} $$ for every fixed $0 0$. For the matching upper-tail classes with a bounded density, these bounds give a full speed-$n$ large deviation principle, with the displayed upper-tail rate for $r\ge1$ and infinite rate for $r<1$. The upper-tail proof uses a change of measure that preserves the entry support and creates an outlying eigenvalue. The lower-tail proof develops a weighted comparison for adaptive orthonormal observations and applies it to Arnoldi residuals.

math.PR↗
arXiv2026

Deployment of AI-Assisted Interventions: Capacity Constraints and Noisy Compliance

AI tools increasingly drive targeted interventions in various service settings, including healthcare, education, and public services. Algorithms score individuals, trigger outreach to those above a threshold (e.g., high-risk or high-value), and encourage them to request service; then providers deliver service to those who request. Much of the work in this area has focused on improving predictive accuracy, implicitly assuming that better predictions lead to better outcomes. We show that predictions are only one component of a larger service system: when service capacity is limited and behavioral responses to outreach are probabilistic, system efficacy depends on operational forces that predictive accuracy does not capture. In such settings, the optimal score threshold must balance two effects: ensuring all capacity is filled (utilization) and, when capacity is constrained, preventing low-value requests from crowding out high-value ones (cannibalization). We characterize the optimal threshold and prove that thresholds based solely on predictive accuracy are generally suboptimal. Further, algorithm selection metrics such as AUC can be misaligned with operational performance: they weight all thresholds equally, while optimal deployment uses a subset of thresholds that depends on both capacity and compliance behavior. We introduce a new metric, Operational AUC (OpAUC), and show that it identifies the efficacy-optimal algorithm. Finally, we conduct a case study on sepsis early warning data that illustrates the magnitude of the improvement available from better threshold selection and shows that a predictor with lower AUC can achieve higher system efficacy under optimal deployment.

stat.ME↗
arXiv2026

Latent Space Is Not Flat: Rethinking Latent Structure for 3D Medical Image Synthesis

Latent generative models make 3D medical image synthesis computationally practical by generating in a compressed space. However, we show that the common flat Euclidean assumption induced by $\ell_2$ objectives is imprecise: latent-space geometry is so strongly anisotropic that equal-magnitude errors can produce drastically different decoded distortions. We further find that this anisotropy has a clear feature: sensitive variation concentrates in a low-rank subspace. The dominant low-rank components capture the overall structure, encoding long-range, spatially coordinated variation while remaining resistant to local noise. Its orthogonal residual, in contrast, mainly captures local and image-specific variation. Motivated by this asymmetry, we introduce Latent Structure Flow (LSF). At each block, LSF decomposes the latent state into structure and residual, models structural changes with global context, and predicts residual variation locally while preserving a direct path for the input structure. LSF changes only the generator, leaving the frozen codec and pointwise training objective unchanged. Across cross-modality synthesis and tumor inpainting tasks, LSF outperforms all compared baselines on both global and tumor-specific metrics, demonstrating the benefit of explicitly modeling latent-space structure for 3D medical image synthesis.

cs.CV↗
arXiv2026

ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation

Active perception and manipulation are crucial for robots to interact with complex scenes. Existing benchmarks struggle to evaluate how robots effectively acquire and maintain information in memory in an active manner. To this end, we introduce ActiveArena-Sim, an active-perception simulator with controllable viewpoints and large-scale workspaces as the foundation. Built on this, we propose ActiveArena-Bench, which comprises 35 tasks across 5 fine-grained categories, covering visual exploration and interactive information acquisition. Each task is difficult to solve from passive observations alone, requiring multi-round evidence acquisition and memory-based reasoning. The benchmark provides rich memory annotations, standardized training data, and ID/OOD protocols featuring disjoint scenes, unseen distractor configurations, and novel backgrounds. Moreover, we present ActiveArena-VLA, a modular suite of 13 vision-language-action configurations for controlled studies of memory writing, memory capacity, proprioceptive state, subtask supervision, and high-level planning in active perception. Benchmark results reveal a substantial ID-OOD gap: uniform memory sampling, increased memory capacity under reliable write policies, proprioceptive inputs, and subtask supervision improve OOD generalization, while planner-guided memory management and decision-making achieve performance close to the best-performing configuration using only sparse memory. ActiveArena thus provides a unified testbed to develop and diagnose models for active perception and manipulation.

cs.RO↗
arXiv2026

LiteSC: Lightweight Semantic Communication for Robust Wireless Telesurgical Video Transmission

Reliable laparoscopic video is essential for telesurgery, yet fixed-rate digital transmission can degrade abruptly under poor channel conditions. We propose LiteSC, a lightweight semantic communication framework combining a frozen pretrained extractor, a compact joint source--channel coding pair, and a receiver-adapted surgical renderer. The extractor produces a latent with approximately 97.9\% fewer scalar values than the red--green--blue (RGB) input, while inference requires neither segmentation masks nor temporal reference frames. Using separate CholecSeg8k source videos for training and testing, we evaluate LiteSC over an additive white Gaussian noise (AWGN) channel at symbol signal-to-noise ratios $E_s/N_0$ from 0 to 25 dB. At a channel bandwidth ratio of $ρ=0.0208$, equivalent to 0.0208 complex channel symbols per RGB source value, peak signal-to-noise ratio (PSNR) rises from 28.7 to 31.9 dB, structural similarity index measure (SSIM) from 0.869 to 0.919, and learned perceptual image patch similarity (LPIPS) falls from 0.102 to 0.048. Against H.265 transmission protected by low-density parity-check (LDPC) coding under the same channel-symbol budget, LiteSC achieves higher SSIM and lower LPIPS, although H.265+LDPC attains a higher peak PSNR. These results show graceful laparoscopic frame reconstruction under the evaluated AWGN conditions.

eess.IV↗
arXiv2026

Outliers and bounded rank perturbation for non-Hermitian random band matrices

In this work we consider general non-Hermitian square random matrices $X$ that include a wide class of random band matrices with independent entries. Whereas the existence of limiting density is largely unknown for these inhomogeneous models, we show that spectral outliers can be determined under very general conditions when perturbed by a finite rank deterministic matrix. More precisely, we show that whenever $\mathbb{E}[X]=0,\mathbb{E}[XX^*]=\mathbb{E}[X^*X]=\mathbf{1}$ and $\mathbb{E}[X^2]=ρ\mathbf{1}$, and under mild conditions on sparsity and entry moments of $X$, then with high possibility all eigenvalues of $X$ are confined in a neighborhood of the support of the elliptic law with parameter $ρ$. Also, a finite rank perturbation property holds: when $X$ is perturbed by another deterministic matrix $C_N$ with bounded rank, then the perturbation induces outlying eigenvalues whose limit depends only on outlying eigenvalues of $C_N$ and $ρ$. This extends the result of Tao on i.i.d. random matrices and O'rourke and Renfrew on elliptic matrices to a family of highly sparse and inhomogeneous random matrices, including all Gaussian band matrices on regular graphs with degree at least $(\log N)^3$. A quantitative convergence rate is also derived. We also consider a class of finite rank deformations of products of at least two independent elliptic random matrices, and show it behaves just as product i.i.d. matrices.

math.PR↗
arXiv2026

The circular law for non-Hermitian random band matrices up to bandwidth $N^{2/3+c}$

We prove circular laws for non-Hermitian random matrices in two regimes. First, for doubly stochastic variance profiles bounded by $C/W$ and satisfying symmetry or transitive invariance, we establish the circular law for real or circular complex Gaussian entries when $W\ge N^{2/3+c}$. No lower variance bound or quantitative mixing assumption is required. A finite-moment comparison extends the result to bounded-density entries at explicit larger bandwidths, including nonperiodic graph supports and inhomogeneous weights. Second, for periodic scalar strips with a flat diagonal core, we prove a least-singular-value estimate without a density assumption and deduce the circular law for real subgaussian atoms, including Rademacher entries, at explicit sublinear power bandwidths. The latter argument combines quantitative Littlewood--Offord counting with a separator-tree elimination and does not require full-block off-diagonal couplings. Both parts use coarse singular-value counting to control logarithmic integrability.

math.PR↗
arXiv2026

The circular law for non-Hermitian random band matrices: optimal bandwidth, periodic profile and discrete law

We consider non-Hermitian random band matrices with growing bandwidth and study convergence of their empirical spectral distributions to the circular law. Let $N$ denote the matrix size and $W_N$ the bandwidth. From a universality perspective, it is conjectured that the circular law holds whenever $W_N\to\infty$. Previous results have mainly required $W_N\gg N^{1/2}$, with the principal exception of \cite{Han2511}, which proves the $W_N\to\infty$ circular law for an open-boundary block-tridiagonal model. Here we prove the circular law at this optimal threshold for several periodic models and under the near-optimal condition $W_N\gg\log N$ for a genuinely discrete model. For the periodic hard-indicator profile and its uniform and polynomially tapered generalizations, we prove the circular law under bounded-density and finite-third-moment assumptions whenever $W_N\to\infty$. At the same threshold, we prove the circular law for continuous, integrable profiles locally bounded below on every finite interval, including exponentially and Gaussian decaying profiles, with circular complex Gaussian entries. For the periodic full-block model, we prove the circular law under bounded-density and finite-third-moment assumptions whenever $W_N\to\infty$. The finite-third-moment assumption in the bounded-density results can be weakened to a finite $(2+α)$-moment assumption. For the same full-block model without a density assumption, we prove the circular law for centered variance-one real subgaussian atoms when $W_N\gg\log N$. The proof uses compositions of random transfer operators. An established high-band circular-law input on a short auxiliary ring calibrates the full exterior coefficient norm, and a boundary-uniform local comparison lifts this calibration to target rings even when $N/W_N$ is arbitrarily large.

math.PR↗