arXiv subjects

Jinpeng Lu

Publications and source records attributed to Jinpeng Lu.

At least 19 recordsLinked to original sources

arXiv2026

VLA-Trace: Diagnosing Vision-Language-Action Models through Representation and Behavior Tracing

Understanding how Vision-Language-Action (VLA) models transform multimodal knowledge into embodied control remains an open challenge. We present VLA-Trace, a progressive diagnostic framework that analyzes VLA models through a unified evidence chain from representation dynamics to causal control attribution and behavioral manifestation. It specifically combines cross-modal and checkpoint-drift centered kernel alignment (CKA) to trace representation evolution, attention knockout interventions to identify modality-specific control pathways, and rollout-level behavioral probes to examine grounding, shortcut dependence, and semantic following. Experiments on $π_{0.5}$ and OpenVLA reveal three key findings. First, the two models exhibit distinct modality-specific adaptation dynamics during VLA finetuning. Second, they rely on different multimodal routing strategies and layer-wise dependencies during action decoding. Third, although VLA policies excel at visually grounded trajectory generation, they remain limited in fine-grained semantic following. These findings highlight future directions for representation-preserving adaptation, causal VLA circuits, and compositional semantic control.

cs.AI↗

arXiv2026

An inverse free boundary problem

We study inverse problems for the elliptic and parabolic obstacle problems from boundary measurements. For the classical elliptic obstacle problem with strictly superharmonic obstacle function, we show that the Dirichlet-to-Neumann map admits a one-sided linearization at every boundary datum lying strictly above the obstacle. The linearized map is the Dirichlet-to-Neumann map for a rough Dirichlet problem on the a priori unknown non-contact set. We show that these linearized Cauchy data uniquely determine the non-contact set up to set of Sobolev $2$-capacity zero and consequently determine the obstacle. Our result applies to the inverse problem for a parabolic obstacle problem where both the coefficient and the obstacle function are time-independent by reducing to the elliptic inverse problem.

math.AP↗

arXiv2026

Johnson-Lindenstrauss Lemma Guided Network for Efficient 3D Medical Segmentation

Lightweight 3D medical image segmentation remains constrained by a fundamental \textit{``efficiency / robustness conflict''}, particularly when processing complex anatomical structures and heterogeneous modalities. In this paper, we study how to redesign the framework based on the characteristics of high-dimensional 3D images, and explore data synergy to overcome the fragile representation of lightweight methods. Our approach, VeloxSeg, begins with a deployable and extensible dual-stream CNN-Transformer architecture composed of Paired Window Attention (PWA) and Johnson-Lindenstrauss lemma-guided convolution (JLC). For each 3D image, we invoke a ``glance-and-focus'' principle, where PWA rapidly retrieves multi-scale information, and JLC ensures robust local feature extraction with minimal parameters, significantly enhancing the model's ability to operate with low computational budget. Followed by an extension of the dual-stream architecture that incorporates modal interaction into the multi-scale image-retrieval process, VeloxSeg efficiently models heterogeneous modalities. Finally, Spatially Decoupled Knowledge Transfer (SDKT) via Gram matrices injects the texture prior extracted by a self-supervised network into the segmentation network, yielding stronger representations than baselines at no extra inference cost. Experimental results on multimodal benchmarks show that VeloxSeg achieves a 26\% Dice improvement, alongside increasing GPU throughput by 11$\times$, CPU by 48$\times$, and reducing training peak GPU memory usage by $1/20$, inference by $1/24$. Code is available at https://github.com/JinPLu/VeloxSeg.

cs.CV↗

arXiv2026

Current World Models Lack a Persistent State Core

World models are increasingly regarded as a decisive step toward artificial general intelligence, yet modeling the physical world demands more than rendering convincing frames on demand: it requires an internal world state that keeps evolving over time, decoupled from observation, so that objects endure and events run to their conclusions whether or not a camera is watching, much as the moon holds to its orbit when no one is looking. This requirement is a blind spot of existing benchmarks, which reward surface properties such as fidelity, motion, and camera controllability while never asking whether a generated world keeps evolving once it is unobserved. We introduce \textbf{WRBench}, the first systematic diagnostic benchmark that treats camera motion as an intervention on observability and resolves evaluation into a human-calibrated chain that asks whether the camera executes the requested interaction, whether the scene stays continuous and identifiable while in view, and whether a returning target remains consistent with the event that was set in motion. Across 9{,}600 videos from 23 models spanning four control paradigms, one finding proves stubborn: current systems maintain the observed world as a tracking shot, resuming a returning target in the state at which it was abandoned rather than advancing the event while it went unseen. Because this failure recurs across control paradigms, model families, and increments of scale, robust world-state evolution does not follow from cleaner imagery, tighter control, richer geometric priors, or sheer parameter count We therefore argue that the stability of the physical state kernel and the consistency of worldlines under viewpoint intervention should become first-class objectives of world-model design, so that a world model captures how the world will unfold rather than how the next frame appears.

cs.CV↗

arXiv2026

Pelican-Unify 1.0: A Unified Embodied Intelligence Model for Understanding, Reasoning, Imagination and Action

We present Pelican-Unify 1.0, the first embodied foundation model trained according to the principle of unification. Pelican-Unify 1.0 uses a single VLM as a unified understanding module, mapping scenes, instructions, visual contexts, and action histories into a shared semantic space. The same VLM also serves as a unified reasoning module, autoregressively producing task-, action-, and future-oriented chains of thought in a single forward pass and projecting the final hidden state into a dense latent variable. A Unified Future Generator (UFG) then conditions on this latent variable and jointly generates future videos and future actions through two modality-specific output heads within the same denoising process. The language, video, and action losses are all backpropagated into the shared representation, enabling the model to jointly optimize understanding, reasoning, imagination, and action during training, rather than training three isolated expert systems. Experiments demonstrate that unification does not imply compromise. With a single checkpoint, Pelican-Unify 1.0 achieves strong performance across all three capabilities: 64.7 on eight VLM benchmarks, the best among comparable-scale models; 66.03 on WorldArena, ranking first; and 93.5 on RoboTwin, the second-best average among compared action methods. These results show that the unified paradigm succeeds in preserving specialist strength while bringing understanding, reasoning, imagination, and action into one model.

cs.RO↗

arXiv2026

An inverse problem for semilinear wave equations on metric tree graphs

We study the inverse problem for a semilinear wave equation on metric tree graphs. From the Dirichlet-to-Neumann map defined at all but one of the boundary vertices, we recover unknown connectivity of the graph, lengths of the edges, the time-independent potential and the time-dependent coefficient of the nonlinear term of the equation.

math.AP↗

arXiv2026

Gel'fand's inverse problem under Ricci curvature bounds

The classical Gel'fand's inverse problem asks whether a Riemannian manifold is uniquely determined by the knowledge of the heat kernel on any open subset of the manifold. We study this inverse problem in the non-smooth setting in the framework of ${\rm RCD}(K,N)$ spaces, namely, metric-measure spaces with synthetic Riemannian Ricci curvature bounded below by $K$ and dimension bounded above by $N$. We establish the unique solvability of Gel'fand's inverse problem for the class of compact ${\rm RCD}(K,N)$ spaces whose regular set admits $C^1$-Riemannian structure. As an application, we obtain the stability of Gel'fand's inverse problem in the class of closed Riemannian manifolds with bounded Ricci curvature, diameter and volume bounded from below. We note that the results are new even for Einstein orbifolds and (weighted) Riemannian manifolds with non-smooth boundary.

math.DG↗

arXiv2026

Does DINOv3 Set a New Medical Vision Standard? Benchmarking 2D and 3D Classification, Segmentation, and Registration

The advent of large-scale vision foundation models, pre-trained on diverse natural images, has marked a paradigm shift in computer vision. However, how the frontier vision foundation models' efficacies transfer to specialised domains such as medical imaging remains an open question. This report investigates whether DINOv3, a state-of-the-art self-supervised vision transformer (ViT) pre-trained on natural images, can directly serve as a powerful, unified encoder for medical vision tasks without domain-specific fine-tuning. To answer this, we benchmark DINOv3 across common medical vision tasks, including 2D and 3D classification, segmentation, and registration on a wide range of medical imaging modalities. We systematically analyse its scalability by varying model sizes and input image resolutions. Our findings reveal that DINOv3 shows impressive performance and establishes a formidable new baseline. Remarkably, it can even outperform medical-specific foundation models like BiomedCLIP and CT-Net on several tasks, despite being trained solely on natural images. However, we identify clear limitations: The model's features degrade in scenarios requiring deep domain specialisation, such as in whole-slide images (WSIs), electron microscopy (EM), and positron emission tomography (PET). Furthermore, we observe that DINOv3 does not consistently follow the scaling law in the medical domain. Its performance does not reliably increase with larger models or finer feature resolutions, showing diverse scaling behaviours across tasks. Overall, our work establishes DINOv3 as a strong baseline, whose powerful visual features can serve as a robust prior for multiple medical tasks. This opens promising future directions, such as leveraging its features to enforce multiview consistency in 3D reconstruction.

cs.CV↗

arXiv2025

IPGPhormer: Interpretable Pathology Graph-Transformer for Survival Analysis

Pathological images play an essential role in cancer prognosis, while survival analysis, which integrates computational techniques, can predict critical clinical events such as patient mortality or disease recurrence from whole-slide images (WSIs). Recent advancements in multiple instance learning have significantly improved the efficiency of survival analysis. However, existing methods often struggle to balance the modeling of long-range spatial relationships with local contextual dependencies and typically lack inherent interpretability, limiting their clinical utility. To address these challenges, we propose the Interpretable Pathology Graph-Transformer (IPGPhormer), a novel framework that captures the characteristics of the tumor microenvironment and models their spatial dependencies across the tissue. IPGPhormer uniquely provides interpretability at both tissue and cellular levels without requiring post-hoc manual annotations, enabling detailed analyses of individual WSIs and cross-cohort assessments. Comprehensive evaluations on four public benchmark datasets demonstrate that IPGPhormer outperforms state-of-the-art methods in both predictive accuracy and interpretability. In summary, our method, IPGPhormer, offers a promising tool for cancer prognosis assessment, paving the way for more reliable and interpretable decision-support systems in pathology. The code is publicly available at https://anonymous.4open.science/r/IPGPhormer-6EEB.

cs.CV↗

arXiv2025

Stability of Gel'fand's inverse interior spectral problem for Schrödinger operators

We study Gel'fand's inverse interior spectral problem of determining a closed Riemannian manifold $(M,g)$ and a potential function $q$ from the knowledge of the eigenvalues $λ_j$ of the Schrödinger operator $-Δ_g + q$ and the restriction of the eigenfunctions $ϕ_j|_U$ on a given open subset $U\subset M$, where $Δ_g$ is the Laplace-Beltrami operator on $(M,g)$. We prove that an approximation of finitely many spectral data on $U$ determines a finite metric space that is close to $(M,g)$ in the Gromov-Hausdorff topology, and further determines a discrete function that approximates the potential $q$ with uniform estimates. This leads to a quantitative stability estimate for the inverse interior spectral problem for Schrödinger operators in the general case.

math.AP↗

arXiv2025

Reconstruction and interpolation of manifolds II: Inverse problems with partial data for distances observations and for the heat kernel

We consider how a closed Riemannian manifold $M$ and its metric tensor $g$ can be approximately reconstructed from local distance measurements. Moreover, we consider an inverse problem of determining $(M,g)$ from limited knowledge on the heat kernel. In the part 1 of the paper, we considered the approximate construction of a smooth manifold in the case when one is given the noisy distances $\tilde d(x,y)=d(x,y)+\varepsilon_{x,y}$ for all points $x,y\in X$, where $X$ is a $δ$-dense subset of $M$ and $|\varepsilon_{x,y}|<δ$. In this part 2 of the paper, we consider a similar problem with partial data, that is, the approximate construction of the manifold $(M,g)$ when we are given $\tilde d(x,y)$ for $x\in X$ and $y \in U\cap X$, where $U$ is an open subset of $M$. In addition, we consider the inverse problem of determining the manifold $(M,g)$ with non-negative Ricci curvature from noisy observations of the heat kernel $G(y,z,t)$. We show that a manifold approximating $(M,g)$ can be determined in a stable way, when for some unknown source points $z_j$ in $X\setminus U$, we are given the values of the heat kernel $G(y,z_k,t)$ for $y\in X\cap U$ and $t\in (0,1)$ with a multiplicative noise. We also give a uniqueness result for the inverse problem in the case when the data does not contain noise and consider applications in manifold learning. A novel feature of the inverse problem for the heat kernel is that the set $M\setminus U$ containing the sources and the observation set $U$ are disjoint.

math.DG↗

arXiv2025

An inverse Signorini obstacle problem

We study the inverse problem of determining a Signorini obstacle from boundary measurements for the isotropic elasticity system. We prove that the obstacle can be uniquely determined by a single measurement of displacement and normal stress for the Signorini problem on an open subset of the boundary up to a natural obstruction. In addition to considering the Signorini problem, we develop techniques that can be used to study inverse problems for general differential inequalities.

math.AP↗

arXiv2024

Stable recovery of coefficients in an inverse fault friction problem

We consider the inverse fault friction problem of determining the friction coefficient in the Tresca friction model, which can be formulated as an inverse problem for differential inequalities. We show that the measurements of elastic waves during a rupture uniquely determine the friction coefficient at the rupture surface with explicit stability estimates.

math.AP↗

arXiv2024

Inverse Spectral Problems for Collapsing Manifolds II: Quantitative Stability of Reconstruction for Orbifolds

We consider the inverse problem of determining the metric-measure structure of collapsing manifolds from local measurements of spectral data. In the part I of the paper, we proved the uniqueness of the inverse problem and a continuity result for the stability in the closure of Riemannian manifolds with bounded diameter and sectional curvature in the measured Gromov-Hausdorff topology. In this paper we show that when the collapse of dimension is $1$-dimensional, it is possible to obtain quantitative stability of the inverse problem for Riemannian orbifolds. The proof is based on an improved version of the quantitative unique continuation for the wave operator on Riemannian manifolds by removing assumptions on the covariant derivatives of the curvature tensor.

math.AP↗

arXiv2024

Inverse Spectral Problems for Collapsing Manifolds I: Uniqueness and Stability

We consider the geometric inverse problem of determining a closed Riemannian manifold from measurements of the heat kernel in an open subset of the manifold. In this paper we analyze the stability of this problem in the class of $n$-dimensional Riemannian manifolds with bounded diameter and sectional curvature. It is well-known that a sequence in this class of manifolds can collapse to a lower dimensional stratified space when the injectivity radius of the sequence of manifolds goes to zero. We prove the uniqueness of the inverse problem on the limiting spaces of the collapsing manifolds. As a result, we obtain stability results for the inverse problem in the class of manifolds with bounded diameter and sectional curvature.

math.DG↗

arXiv2024

Rethinking Attention-Based Multiple Instance Learning for Whole-Slide Pathological Image Classification: An Instance Attribute Viewpoint

Multiple instance learning (MIL) is a robust paradigm for whole-slide pathological image (WSI) analysis, processing gigapixel-resolution images with slide-level labels. As pioneering efforts, attention-based MIL (ABMIL) and its variants are increasingly becoming popular due to the characteristics of simultaneously handling clinical diagnosis and tumor localization. However, the attention mechanism exhibits limitations in discriminating between instances, which often misclassifies tissues and potentially impairs MIL performance. This paper proposes an Attribute-Driven MIL (AttriMIL) framework to address these issues. Concretely, we dissect the calculation process of ABMIL and present an attribute scoring mechanism that measures the contribution of each instance to bag prediction effectively, quantifying instance attributes. Based on attribute quantification, we develop a spatial attribute constraint and an attribute ranking constraint to model instance correlations within and across slides, respectively. These constraints encourage the network to capture the spatial correlation and semantic similarity of instances, improving the ability of AttriMIL to distinguish tissue types and identify challenging instances. Additionally, AttriMIL employs a histopathology adaptive backbone that maximizes the pre-trained model's feature extraction capability for collecting pathological features. Extensive experiments on three public benchmarks demonstrate that our AttriMIL outperforms existing state-of-the-art frameworks across multiple evaluation metrics. The implementation code is available at https://github.com/MedCAI/AttriMIL.

cs.CV↗

arXiv2024

H2ASeg: Hierarchical Adaptive Interaction and Weighting Network for Tumor Segmentation in PET/CT Images

Positron emission tomography (PET) combined with computed tomography (CT) imaging is routinely used in cancer diagnosis and prognosis by providing complementary information. Automatically segmenting tumors in PET/CT images can significantly improve examination efficiency. Traditional multi-modal segmentation solutions mainly rely on concatenation operations for modality fusion, which fail to effectively model the non-linear dependencies between PET and CT modalities. Recent studies have investigated various approaches to optimize the fusion of modality-specific features for enhancing joint representations. However, modality-specific encoders used in these methods operate independently, inadequately leveraging the synergistic relationships inherent in PET and CT modalities, for example, the complementarity between semantics and structure. To address these issues, we propose a Hierarchical Adaptive Interaction and Weighting Network termed H2ASeg to explore the intrinsic cross-modal correlations and transfer potential complementary information. Specifically, we design a Modality-Cooperative Spatial Attention (MCSA) module that performs intra- and inter-modal interactions globally and locally. Additionally, a Target-Aware Modality Weighting (TAMW) module is developed to highlight tumor-related features within multi-modal features, thereby refining tumor segmentation. By embedding these modules across different layers, H2ASeg can hierarchically model cross-modal correlations, enabling a nuanced understanding of both semantic and structural tumor features. Extensive experiments demonstrate the superiority of H2ASeg, outperforming state-of-the-art methods on AutoPet-II and Hecktor2022 benchmarks. The code is released at https://github.com/JinPLu/H2ASeg.

eess.IV↗

arXiv2024

Quantitative stability of Gel'fand's inverse boundary problem

In Gel'fand's inverse problem, one aims to determine the topology, differential structure and Riemannian metric of a compact manifold $M$ with boundary from the knowledge of the boundary $\partial M,$ the Neumann eigenvalues $λ_j$ and the boundary values of the eigenfunctions $φ_j|_{\partial M}$. We show that this problem has a stable solution with quantitative stability estimates in a class of manifolds with bounded geometry. More precisely, we show that finitely many eigenvalues and the boundary values of corresponding eigenfunctions, known up to small errors, determine a metric space that is close to the manifold in the Gromov-Hausdorff sense. We provide an algorithm to construct this metric space. This result is based on an explicit estimate on the stability of the unique continuation for the wave operator.

math.AP↗