arXiv subjects

Xiaojie Wang

Publications and source records attributed to Xiaojie Wang.

At least 19 recordsLinked to original sources

arXiv2026

Enhanced convergence guarantees of score-based generative models in $\mathcal{W}_2$-distance beyond log-concavity

Although score-based generative models (SGMs) have achieved remarkable success in real-world sample generation tasks, it is still far from sufficient to understand them from a mathematical perspective. In this paper, we aim to provide enhanced convergence guarantees for two existing SGMs in $\mathcal{W}_2$-distance beyond log-concavity. More precisely, we develop a novel framework of {\color{black}discretization} error analysis for Euler-Maruyama (EM) and Poisson midpoint (PM) time discretization schemes for SGMs. Under a non-log-concavity condition, we show that $\tilde{\mathcal{O}}(\sqrt{d}/ε)$ iterations suffice to approximate the target distributions in $ε$-accuracy for the classical EM-based SGM, considerably improving upon the existing iteration complexity $\tilde{\mathcal{O}}(d/ε^2)$. Notably, the novel framework of error analysis also enables us to establish an $\tilde{\mathcal{O}}(\sqrt{d}/ε^{2/3})$ iteration complexity for the PM-based SGM in $ε$-accuracy, significantly outperforming current state-of-the-art Wasserstein convergence guarantees for SDE-based diffusion samplers.

math.ST↗

arXiv2026

HP-PnP-Flow: Hybrid Time Scheduling and Projection for Plug-and-Play Image Restoration via Flow Matching

Image restoration is a fundamental inverse problem that seeks to recover high-quality images from degraded observations. Plug-and-Play (PnP) methods provide a flexible framework for integrating pretrained denoisers into iterative reconstruction, whereas flow matching offers an efficient generative modeling approach based on continuous-time ODE dynamics. Very recently, Martin et al.\ (ICLR, 2025) developed a PnP-Flow method, which integrates the PnP framework with flow matching through a time-dependent denoiser constructed from a pretrained flow matching model.By incorporating a different time schedule and an additional pixel-space projection operator, this paper proposes a novel efficient PnP-Flow algorithm (HP-PnP-Flow) that enjoys rigorous convergence guarantees under mild assumptions, without imposing restrictive conditions on iterates required by the aforementioned work. Numerical experiments on CelebA and AFHQ-Cat demonstrate that the newly proposed HP-PnP-Flow algorithm significantly outperforms PnP-Flow on several restoration tasks such as denoising, deblurring, super-resolution, random inpainting, and box inpainting. Notably, our algorithm achieves comparable or better reconstruction quality with substantially fewer iterations, reducing the iteration count from $500$ to $100$ for deblurring and super-resolution on AFHQ-Cat.

cs.CV↗

arXiv2026

Sample-Efficient Reinforcement Learning for Policy Improvement via Flow-Reversed Action Anchors

Generalist robot policies offer strong behavioral priors, but achieving deployment-level reliability and throughput remains difficult. Reinforcement learning (RL) can improve these policies, yet its effectiveness is constrained by the cost of online interaction. Existing methods adapt frozen flow-matching policies by treating noise space as a continuous action space. We show that this behaviorally unstructured space makes exploration inefficient, often preventing these methods from extracting meaningful learning signals on challenging tasks. We introduce AnchorQ, which enables coarse-to-fine policy improvement through a structured, refinable latent action representation. At each decision step, AnchorQ uses flow reversal to map reference behaviors to state-conditioned latent anchors, samples local particles around those anchors, and applies value-guided sampling to explore promising refinements. Our analysis shows that anchors reduce identification complexity at the cost of support bias; particle refinement trades this bias against action-class complexity; and critic accuracy determines the reliability of KL-regularized, value-guided selection. Across simulated and real-world manipulation tasks, AnchorQ outperforms state-of-the-art post-training baselines in sample efficiency, delivering a 37% improvement in real-world success rate and 97% higher simulated task throughput under the same interaction budget. Additional visualizations are available on our website https://anchorq-rl.com/.

cs.RO↗

arXiv2026

Self-Refining Vision Language Model for Robotic Failure Detection and Reasoning

Reasoning about failures is crucial for building reliable and trustworthy robotic systems. Prior approaches either treat failure reasoning as a closed-set classification problem or assume access to ample human annotations. Failures in the real world are typically subtle, combinatorial, and difficult to enumerate, whereas rich reasoning labels are expensive to acquire. We address this problem by introducing ARMOR: Adaptive Round-based Multi-task mOdel for Robotic failure detection and reasoning. We formulate detection and reasoning as a multi-task self-refinement process, where the model iteratively predicts detection outcomes and natural language reasoning conditioned on past outputs. During training, ARMOR learns from heterogeneous supervision - large-scale sparse binary labels and small-scale rich reasoning annotations - optimized via a combination of offline and online imitation learning. At inference time, ARMOR generates multiple refinement trajectories and selects the most confident prediction via a self-certainty metric. Experiments across diverse environments show that ARMOR achieves state-of-the-art performance by improving over the previous approaches by up to 30% on failure detection rate and up to 100% in reasoning measured through LLM fuzzy match score, demonstrating robustness to heterogeneous supervision and open-ended reasoning beyond predefined failure modes. We provide additional visualizations on our website: https://sites.google.com/utexas.edu/armor

cs.RO↗

arXiv2026

GSM: Efficient Language Modeling with Shared Global State

Improving language model efficiency requires not only cheaper operations but also eliminating redundant computation. We argue that effective modeling requires information accessibility and computational depth, but not repeated historical retrieval across layers. Guided by this principle, we introduce the Global State Model (GSM), a causal encoder--decoder architecture that separates historical information acquisition from deep computation. The encoder progressively aggregates long-range information through multi-stage sparse retrieval into a fixed-size shared state, while decoder layers perform successive transformations over this state without repeated historical access. This design makes decoder attention and cache costs independent of history length, while dependency-aware pruning further reduces prefill computation. Experiments across multiple model scales show that GSM achieves downstream performance comparable to matched DeepSeek-V4.1-Flash-based baselines, with improvements at some scales, while preserving long-context retrieval capabilities. GSM also improves training and inference throughput and reduces cache overhead. These results demonstrate that eliminating architectural redundancy offers an effective path toward efficient language modeling without sacrificing information accessibility or computational depth.

cs.CL↗

arXiv2026

DT-LoRA: Reaching Beyond the Tangent Space of Low-Rank Adaptation

Low-rank adaptation (LoRA) often underperforms full fine-tuning. We investigate this performance gap through a tangent-space lens: factorized updates are confined to the tangent space of the current low-rank factors, leaving gradient components outside this space unexploited. An analysis of ten representative LoRA variants reveals that these residual components account for most of the gradient energy and exhibit strong spectral concentration. Motivated by this finding, we propose Dynamic Tangent LoRA (DT-LoRA), which partitions a fixed rank budget into a stable branch and a dynamic branch to capture the principal residual gradient directions. We show that these directions define the optimal update of a local quadratic model under orthogonal-complement and rank constraints. Matrix-free randomized SVD recovers them without materializing the residual gradient, while controlled write-back preserves accumulated updates and stabilizes training within the allocated rank. DT-LoRA adds no trainable parameters or inference cost over LoRA. Experiments across computer vision and natural language processing tasks demonstrate consistent improvements over LoRA and its variants. Moreover, despite being a general-purpose adaptation method, DT-LoRA outperforms task-specific approaches to open-vocabulary domain generalization for semantic segmentation, achieving state-of-the-art results. Code is available at https://github.com/Snipine/DT-LoRA.

cs.CV↗

arXiv2026

Generalizable Dense Reward for Long-Horizon Robotic Tasks

Existing robotic foundation policies are trained primarily via large-scale imitation learning. While such models demonstrate strong capabilities, they often struggle with long-horizon tasks due to distribution shift and error accumulation. While reinforcement learning (RL) can finetune these models, it cannot work well across diverse tasks without manual reward engineering. We propose VLLR, a dense reward framework combining (1) an extrinsic reward from Large Language Models (LLMs) and Vision-Language Models (VLMs) for task progress recognition, and (2) an intrinsic reward based on policy self-certainty. VLLR uses LLMs to decompose tasks into verifiable subtasks and then VLMs to estimate progress to initialize the value function for a brief warm-up phase, avoiding prohibitive inference cost during full training; and self-certainty provides per-step intrinsic guidance throughout PPO finetuning. Ablation studies reveal complementary benefits: VLM-based value initialization primarily improves task completion efficiency, while self-certainty primarily enhances success rates, particularly on out-of-distribution tasks. On the CHORES benchmark covering mobile manipulation and navigation, VLLR achieves up to 56% absolute success rate gains over the pretrained policy, up to 5% gains over state-of-the-art RL finetuning methods on in-distribution tasks, and up to $10\%$ gains on out-of-distribution tasks, all without manual reward engineering. Additional visualizations can be found in https://silongyong.github.io/vllr_project_page/

cs.RO↗

arXiv2026

XMM-Newton Resolves Parsec-scale X-ray Jets in the PeVatron Microquasar V4641 Sgr

The microquasar V4641 Sagittarii (V4641 Sgr) has recently been established as one of the most powerful Galactic PeV cosmic-ray accelerators (PeVatrons) through the detection of degree-scale gamma-ray emission to 0.8 PeV. Using XMM-Newton, we performed an X-ray survey of its gamma-ray-emitting regions during the quiescent phase. We discovered extended X-ray emission to the north and south of V4641 Sgr, aligned with the radio-jet axis, revealing X-ray jet structures extending over several tens of parsecs for the first time. The northern and southern knot-like features, dubbed n1 and s1, are located approximately 20 pc from the central binary and spatially coincident with the gamma-ray emission. The X-ray emission from each knot, characterized by a photon index of $Γ_X = 1.9 \pm 0.1_{\rm stat} \pm 0.5_{\rm syst}$ and an energy flux of $\sim$(2--4)$\times10^{-13}$~erg~cm$^{-2}$~s$^{-1}$ in the 2--10 keV band, is likely non-thermal in origin, indicating synchrotron radiation from accelerated electrons. The morphological coincidence between the X-ray and gamma-ray emission favors a leptonic scenario for the gamma rays, although the X-ray emission is more compact than the gamma-ray-emitting region and the hadronic scenario cannot be excluded. Through X-ray to gamma-ray modeling, we obtain a weak magnetic field strength of 3--5~$μ$G. We propose that the newly detected X-ray knots represent particle-acceleration sites, such as jet termination or (re-)collimation regions powered by V4641 Sgr, or downstream structures analogous to the knots of the outer jets in SS~433/W50. This discovery suggests that extended X-ray emission along jets would be a key signature, yet previously overlooked, in PeVatron microquasars.

astro-ph.HE↗

arXiv2026

Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations

Transformers lack a native lookup mechanism, requiring repeated dense computation to recognize and reuse local static patterns. Lngram v1 introduces tokenizer-independent conditional memory through discrete latent n-gram addressing, but its memory capacity is coupled with the backbone width, limiting scalability due to high parameter and activation costs. We propose Lngram v2, which decouples the number of routes, memory dimension, and backbone width, and introduces a context-aware grouped-query attention readout to scale memory capacity independently. A zero-value Sink and counterfactual surrogate gradients further improve readout selectivity and routing trainability while preserving hard discrete addressing. Experiments across vision--language models (VLMs) of different scales show consistent improvements, including successful scaling to a 30B-parameter model. Compared with Lngram v1, Lngram v2 substantially reduces both total and activated memory parameters while maintaining or improving language modeling performance. Further analysis shows that its discrete IDs preserve substantial semantic structure of continuous hidden states, enabling semantic recovery from IDs alone and stable ID--semantic associations across datasets. These results establish Lngram v2 as an efficient and scalable latent conditional memory mechanism whose discrete addresses also provide a structured interface for analyzing internal model representations.

cs.CL↗

arXiv2026

Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms

Despite impressive visual quality, state-of-the-art video diffusion models often generate content that violates real-world physical laws. While existing solutions rely on external priors or specialized data, we investigate the root cause by exploring the internal mechanisms of these models. Specifically, we present the first interpretability study on the ''motion planning'' process of text-to-video diffusion models, revealing how motion trajectories form during early denoising stages. Building upon the ''first shape, then details'' finding, we combine cross-attention trajectory patterns with causal head contributions to identify a specific subset of attention heads driving motion planning. Further, our self-attention analysis shows that Rotary Position Embedding (RoPE) induces excessive spatial attention decay. This causes early candidate regions to prematurely lock into physically implausible positions, suppressing reasonable trajectories in adjacent frames and triggering generation failure modes. To address this fundamental flaw, we propose a lightweight architectural modification that scales the frequency of RoPE across different denoising steps. This strategy reduces excessive attention decay, helping the model explore better candidate regions to establish coherent physical motion. Finally, training-free and training-based experiments confirm the effectiveness of our approach in enhancing the physical commonsense of generated videos.

cs.CV↗

arXiv2026

A Unified Framework for Wasserstein Convergence of ULMC Methods beyond Log-Concavity: Old and New

As a fundamental task across computational statistics, scientific computing and machine learning, sampling from high-dimensional probability distributions has received increasing attention in recent years. Numerous sampling algorithms have been proposed, among which underdamped Langevin Monte Carlo (ULMC) methods based on underdamped Langevin dynamics (ULD) have emerged as a class of efficient ones. In this work, we introduce a ``universal" predictor-corrector formulation that bridges Euler-type, UBU-type and randomized schemes through different choices of method parameters. Notably, the ``universal" integrator induces two novel classes of low-cost integrators, termed low-cost randomized integrators (LC-RIs) and low-cost UBU integrators (LC-UBUIs), as well as their exponential-free variants based on polynomial and rational approximations. The resulting new UBU-type and randomized schemes require only one gradient evaluation and two Gaussians per iteration, considerably reducing the number of gradient evaluations or Gaussians per iteration required by existing counterparts. Further, a general framework of long-time error analysis is developed for general discretization schemes in a probability metric. Under certain smoothness and non-log-concavity conditions, we rely on the unified framework to establish non-asymptotic $\mathcal{W}_1$-error bounds of both old and new schemes, revealing convergence rates of order $\mathcal{O}(d^{\frac{1}{2}}h)$ for Euler-type schemes, order $\mathcal{O}(d h^2)$ for UBU-type ones and order $\mathcal{O}(d^{\frac{1}{2}}h^{\frac{3}{2}})$ for randomized ones. In the strongly convex setting, the same non-asymptotic error bounds can be recovered in $\mathcal{W}_2$-distance. Numerical experiments corroborate the theoretical findings.

math.NA↗

arXiv2026

Multi-Touch and Bending Sensing Using Electrical Impedance Tomography for Robotics

Electrical Impedance Tomography (EIT) offers a promising solution for distributed tactile sensing with minimal wiring and full-surface coverage in robotic applications. However, EIT-based tactile sensors face significant challenges during surface bending. Deformation alters the baseline impedance distribution and couples with touch-induced conductivity variations, complicating signal interpretation. To address this challenge, we present a novel sensing framework that integrates a deep neural network for interaction state classification with a dynamic adaptive reference strategy to decouple touch and deformation signals, while a data-driven regression model translates EIT voltage changes into continuous bending angles. The framework is validated using a magnetic hydrogel composite sensor that conforms to bendable surfaces. Experimental evaluations demonstrate that the proposed framework achieves precise and robust bending angle estimation, high accuracy in distinguishing touch, bending, and idle states, and significantly improves touch localization quality under bending deformation compared to conventional fixed-reference methods. Real-time experiments confirm the system's capability to reliably detect multi-touch interactions and track bending angles across varying deformation conditions. This work paves the way for flexible EIT-based robotic skins capable of rich multimodal sensing in robotics and human-robot interaction.

cs.RO↗

arXiv2026

Integrated polarization-entangled photon source for wavelength-multiplexed quantum networks

Entangled photons are fundamental resources for quantum communication, computing, and networking. Among them, polarization-entangled photon pairs play an important role due to their straightforward state manipulation and direct use in quantum key distribution, teleportation, and network protocols. However, realizing compact, efficient, and scalable polarization-entangled sources that meet the requirements of practical deployment remains a major challenge. Here, we present a simple yet high-performance on-chip polarization-entangled photon-pair source on thin-film lithium niobate (TFLN). Our device employs dual quasi-phase matching (D-QPM) that sequentially supports type-0 and type-I spontaneous parametric down-conversion in a single nanophotonic waveguide, eliminating the need for interferometers, polarization rotators, or other complex circuits. The source directly produces high-fidelity Bell states with broad bandwidth, high brightness, and low noise. Using this integrated platform, we realize wavelength-multiplexed entanglement distribution in a four-user quantum network deployed over metropolitan fiber links up to 50 km. These results establish a robust and scalable pathway toward practical quantum communication systems and multi-user quantum mesh networks based on integrated photonics.

physics.optics↗

arXiv2026

Enhancing Multilingual Reasoning via Steerable Model Merging

Model merging is an effective technique for composing the capabilities of a multilingual model and a reasoning model. It has achieved promising generalization in multilingual reasoning tasks by aligning feature spaces of different models. However, the merged single model often fails to address the conflicts between source models, leading to suboptimal performance. In other words, the one-size-fits-all merging strategy may not align with the characteristics of different inputs which may require prioritizing certain models over others. To this end, we propose a Steerable Model Merging (ST-Merge) framework to modulate the contribution of each source model. To realize this idea, we introduce a gated cross-attention mechanism to weight or filter the two attended source models in an adaptive manner. Extensive experiments demonstrate that ST-Merge consistently outperforms multiple strong baselines on four multilingual reasoning benchmarks across 21 different languages.

cs.CL↗

arXiv2026

PTL-Diffusion: Manifold-Aware Diffusion with Periodic Terminal Laws

Standard diffusion models typically use a single time-homogeneous Gaussian terminal distribution as the reference law for generation. While this choice is analytically convenient and empirically powerful, it provides little explicit structure for data concentrated near low-dimensional manifolds, where different regions of the data distribution may correspond to distinct local geometric or semantic factors. As a result, the reverse model must recover manifold-level structure almost entirely from an unstructured terminal reference distribution. We propose PTL-Diffusion, a proof-of-concept diffusion framework whose forward noising process converges to a nonconstant periodic family of Gaussian terminal laws rather than to a single invariant law. Unlike a phase-conditioned DDPM, where phase information only enters the denoising network while the forward process remains unchanged, PTL-Diffusion embeds phase structure directly into the forward noising dynamics. The proposed construction remains close to standard denoising diffusion models: for a periodically forced Ornstein--Uhlenbeck-type forward process, we derive closed-form forward marginals, the limiting periodic Gaussian terminal family, and explicit Gaussian reverse posteriors, enabling standard noise-prediction training. We also introduce an invariant-average regularization term coupling the phase-conditioned reverse dynamics through the averaged periodic reference law. Experiments on torus and cylinder point-cloud benchmarks and the Olivetti face dataset show that PTL-Diffusion improves manifold-level distributional matching over matched DDPM baselines, reducing phase-conditioned errors, feature-space covariance errors, and nearest-neighbour manifold distances. These results suggest structured terminal reference laws as a promising direction, while motivating more expressive phase constructions and larger-scale evaluations.

cs.CV↗

arXiv2026

Multimodal sampling via Schrödinger-Föllmer samplers with temperatures

Generating samples from complex and high-dimensional distributions is ubiquitous in various scientific fields of statistical physics, Bayesian inference, scientific computing and machine learning. Very recently, Huang et al. (IEEE Trans. Inform. Theory, 2025) proposed new Schrödinger-Föllmer samplers (SFS), based on the Euler discretization of the Schrödinger-Föllmer diffusion evolving on the unit interval $[0, 1]$. There, a convergence rate of order $\mathcal{O}(\sqrt{h})$ in the $L^2$-Wasserstein distance was obtained for the Euler discretization with a uniform time step-size $h>0$. By introducing a temperature parameter, different samplers are proposed in this paper, based on the Euler discretization of the Schrödinger-Föllmer process with temperatures. As revealed by numerical experiments, high temperatures are vital, particularly in sampling from multimodal distributions. Further, a novel approach of error analysis is developed for the time discretization and an enhanced convergence rate of order $\mathcal{O}(h)$ is obtained in the $L^2$-Wasserstein distance, under certain smoothness conditions on the drift. This significantly improves the existing order-half convergence in the aforementioned paper. Unlike Langevin samplers, SFS is gradient-free, works in a unit interval $[0, 1]$ and does not require any ergodicity. Numerical experiments confirm the convergence rate and show that, the SFS substantially outperforms vanilla Langevin samplers, particularly in sampling from multimodal distributions.

math.NA↗

arXiv2026

Accelerated Schrödinger-Föllmer samplers

Sampling is a fundamental algorithmic task in wide-ranging applications across multiple disciplines such as scientific computing, statistics and machine learning. In this paper, an efficient stochastic Runge-Kutta scheme is proposed to accelerate the Schrödinger-Föllmer sampler, designed for sampling from complex and high-dimensional multimodal distributions. The resulting stochastic Runge-Kutta Schrödinger-Föllmer sampler (SRKSFS) is proved to achieve a convergence rate of order $\mathcal{O} ( h^{3/2} |\ln h|)$ in the $L^2$-Wasserstein distance, considerably improving the order $\mathcal{O}(h)$ of the existing Euler type sampler. Obtaining the enhanced convergence rate is, however, not trivial, by noting that the drift of the diffusion process is not differentiable but only $\frac{1}{2}$-Hölder continuity with respect to the time variable. To address the difficulty, we rely on delicate error estimates to overcome the singularity due to time derivatives of the drift, at the expense of the logarithmic factor. Furthermore, the framework is extended to data-driven Schrödinger-Föllmer generation with empirical measures, enabling data-driven sampling without known density. A variety of numerical experiments are reported to validate the effectiveness of the proposed sampling algorithms.

math.ST↗

arXiv2026

ChunkLLM: A Lightweight Pluggable Framework for Accelerating LLMs Inference

Transformer-based large models excel in natural language processing and computer vision, but face severe computational inefficiencies due to the self-attention's quadratic complexity with input tokens. Recently, researchers have proposed a series of methods based on block selection and compression to alleviate this problem, but they either have issues with semantic incompleteness or poor training-inference efficiency. To comprehensively address these challenges, we propose ChunkLLM, a lightweight and pluggable training framework. Specifically, we introduce two components: QK Adapter (Q-Adapter and K-Adapter) and Chunk Adapter. The former is attached to each Transformer layer, serving dual purposes of feature compression and chunk attention acquisition. The latter operates at the bottommost layer of the model, functioning to detect chunk boundaries by leveraging contextual semantic information. During the training phase, the parameters of the backbone remain frozen, with only the QK Adapter and Chunk Adapter undergoing training. Notably, we design an attention distillation method for training the QK Adapter, which enhances the recall rate of key chunks. During the inference phase, chunk selection is triggered exclusively when the current token is detected as a chunk boundary, thereby accelerating model inference. Experimental evaluations are conducted on a diverse set of long-text and short-text benchmark datasets spanning multiple tasks. ChunkLLM not only attains comparable performance on short-text benchmarks but also maintains 98.64% of the performance on long-context benchmarks while preserving a 48.58% key-value cache retention rate. Particularly, ChunkLLM attains a maximum speedup of 4.48x in comparison to the vanilla Transformer in the processing of 120K long texts.

cs.CL↗