arXiv:2603.16011v3 Announce Type: replace
Abstract: Large language model (LLM) coding agents increasingly operate at the repository level, motivating benchmarks that evaluate their ability to optimize entire codebases under realistic constraints. Existing code benchmarks largely rely on synthetic tasks, binary correctness signals, or single-objective evaluation, limiting their ability to assess holistic optimization behavior. We introduce FormulaCode, a benchmark for evaluating agentic optimization on large, real-world codebases with fine-grained, multi-objective performance metrics. FormulaCode comprises 957 performance bottlenecks mined from scientific Python repositories on GitHub, each paired with expert-authored patches and, on average, 264.6 community-maintained performance workloads per task, enabling the holistic ability of LLM agents to optimize codebases under realistic correctness and performance constraints. Our evaluations reveal that repository-scale, multi-objective optimization remains a major challenge for frontier LLM agents. Project website at: https://formula-code.github.io
Science Journals
arXiv:2603.16403v2 Announce Type: replace
Abstract: Digital identity is shifting from service- and network-centric approaches toward user-centric ones that promise users increased control over their data. Despite their decentralised design, such approaches often reintroduce centralised components in different forms. This paper conceptualises this tension as the decentralisation paradox and shows that user-centric architectures tend to redistribute rather than eliminate centralisation. Drawing on Critical Systems Thinking (CST), digital identity is framed as a "wicked problem" that spans the technical, legal, social and ethical dimensions. It introduces the Digital Identity Tetrahedron as a multidimensional framework for analysing decentralisation in digital identity ecosystems. Understanding these interdependencies is essential for designing reliable architectures and ensuring that the next generation of digital identity surpasses superficial decentralisation.
arXiv:2603.18046v2 Announce Type: replace
Abstract: We present NanoZK, a zero-knowledge proof system for verifiable LLM inference: clients and third-party auditors check that a provider executed the advertised model on a committed input without learning weights or activations. NanoZK introduces a layerwise proof framework that decomposes transformer inference into independently provable layers linked by a SHA-256 commitment chain, yielding constant-size sub-circuit proofs (3.5-3.7 KB; about 83 KB total at L=12), comparable in total size to and substantially more parallelizable than prior ZKML's monolithic 101-126 KB proofs. We prove compositional soundness and zero-knowledge under standard assumptions, design 16-bit lookup-table approximations for softmax, GELU, and normalization with measured perplexity degradation below 1e-4 across six model/dataset combinations, and add a Fisher-information-guided audit-budget triage as an efficiency tool (full soundness still requires verifying every layer). On CPU the MLP sub-circuit proves in about 6.3 s prove-only (about 43 s setup plus prove) with about 22 ms verification at any width; attention prove-only time scales from 0.9 s (d=16) to 184 s (d=256); full-block end-to-end proofs are measured to d=128, with a projected GPU time of about 68 s per block at d=768 from measured O(d^2) MSM scaling and a conservative 15-30x GPU-MSM speedup range based on Icicle's published 30x result for n >= 2^20 and extrapolated to the smaller-n regime. Privacy scope: NanoZK hides weights and activations from verifiers and auditors but does not hide the prompt from the prover; this is complementary to HE/MPC.
arXiv:2603.18482v3 Announce Type: replace
Abstract: Why does machine-generated text remain detectable? We trace the answer to the decoding stage: standard strategies such as top-$k$ and nucleus sampling restrict generation to high-probability tokens, while human writers routinely choose words that are contextually apt yet rank low under a model's next-token distribution. Truncation therefore renders a measurable share of human word choices unreachable; we call this set the truncation blind spot. Across five open models and three domains, 8--18\% of human-selected tokens fall outside common truncation boundaries, with content words excluded at 2.9 times the rate of function words. As a consequence, simple classifiers using two features separate 1.8 million machine generations from human text at mean AUC-ROC near 0.97, and detectability tracks truncation intensity rather than model scale, architecture, or alignment. Probability-floor samplers substantially narrow the blind spot, and generation experiments confirm the exclusion--detectability link out of sample. Code, data, and the measurement harness are released at https://github.com/EstebanGarces/human_vs_machine.
arXiv:2603.18937v2 Announce Type: replace
Abstract: In classical information theory, both the form and performance of the optimal detector for additive noise channels can be precisely derived, based on the assumption that the channel noise follows a specific probability distribution or a mixture of known distributions, or that the exact distribution exists but is unknown. In this paper, we extend the analyses of detectors for additive noise channel to the situation where the probability model for analyzing channels is uncertain, utilizing nonlinear expectation theory. We consider two types of distribution uncertainties: one with no mean uncertainty but with variance uncertainty, and another with both mean and variance uncertainties. We derive the optimal threshold detectors for binary input additive noise channel under the nonlinear expectation optimal criterion for both scenarios and provide their explicit forms. Our findings reveal that mean uncertainty significantly influences the form of the optimal detector, whereas variance uncertainty does not. Additionally, we propose an estimation method for the uncertain parameters of the channel noise. Finally, we present theoretical analyses and simulated performance results of the newly derived optimal threshold detectors, and compare these results with the performance of optimal detector under classical information theory, which assumes a deterministic probability model. The results of experiments show that our new detection methods outperform conventional methods in most scenarios with uncertain probability models, showing the practical relevance of our theoretical contributions.
arXiv:2603.20714v3 Announce Type: replace
Abstract: 3D Gaussian Splatting (3DGS) has become the method of choice for photo-realistic novel view synthesis (NVS), due to its efficiency and compelling visual quality. 3DGS represents the scene through a set of 3D Gaussians, parameterized by their position, spatial extent, and view-dependent color. Starting from an initial point cloud, 3DGS refines the Gaussians' parameters as to reconstruct a set of training images as accurately as possible. Typically, a sparse Structure-from-Motion point cloud is used as initialization. In order to obtain a full scene representation, 3DGS methods thus rely on a densification stage. In this paper, we systematically study how initialization affects 3DGS NVS performance and geometric quality, using several densification strategies. We show that dense initialization does not lead to consistent visual improvements when paired with strong densification. Despite that, it can help in generalization to off-trajectory views and significantly improves geometric consistency of the scene representation.
arXiv:2603.20843v3 Announce Type: replace
Abstract: Long-context language modeling is commonly framed as a scalability challenge of token-level attention, yet local-to-global information structuring remains largely implicit in existing approaches. Drawing on cognitive theories of discourse comprehension, we propose HiCI (Hierarchical Construction--Integration), a hierarchical attention module that constructs segment-level representations, integrates them into a shared global context, and broadcasts both to condition segment-level attention. We validate HiCI through parameter-efficient adaptation of LLaMA-2 with only <5.5% additional parameters, extending context from 4K to 100K tokens (7B) and 64K tokens (13B). Across language modeling, retrieval, and instruction-following benchmarks, HiCI yields consistent improvements over strong baselines, including matching proprietary models on topic retrieval and surpassing GPT-3.5-Turbo-16K on code comprehension. These results demonstrate the effectiveness of explicit hierarchical structuring as an inductive bias for long-context modeling.
arXiv:2603.21693v2 Announce Type: replace
Abstract: Multimodal large language models (MLLMs) have shown strong potential for medical Visual Question Answering (VQA), yet they remain prone to hallucinations, defined as generating responses that contradict the input image, posing serious risks in clinical settings. Current hallucination detection methods, such as Semantic Entropy (SE) and Vision-Amplified Semantic Entropy (VASE), require 10 to 20 stochastic generations per sample together with an external natural language inference model for semantic clustering, making them computationally expensive and difficult to deploy in practice. We observe that hallucinated responses exhibit a distinctive signature directly in the model's own log-probabilities: inconsistent token-level confidence and weak sensitivity to visual evidence. Based on this observation, we propose Confidence-Evidence Bayesian Gain (CEBaG), a deterministic hallucination detection method that requires no stochastic sampling, no external models, and no task-specific hyperparameters. CEBaG combines two complementary signals: token-level predictive variance, which captures inconsistent confidence across response tokens, and evidence magnitude, which measures how much the image shifts per-token predictions relative to text-only inference. Evaluated across four medical MLLMs and three VQA benchmarks (16 experimental settings), CEBaG achieves the highest AUC in 13 of 16 settings and improves over VASE by 8 AUC points on average, while being fully deterministic and self-contained. https://github.com/masadi-99/CEBaG
arXiv:2603.21876v3 Announce Type: replace
Abstract: Infrared pedestrian detectors are increasingly deployed in all-weather perception systems, but their robustness against physical adversarial attacks remains insufficiently understood. Existing infrared physical attacks are mostly instance-specific, requiring perturbations to be optimized for particular samples, poses, or scenes, which limits scalability under changing deployment conditions. This paper proposes Universal Physical Patch Attack (UPPA), a universal cold-patch framework for infrared pedestrian detection. UPPA is built on the observation that infrared physical perturbations should exploit smooth, low-frequency thermal structures rather than visible-light texture patterns. It represents the attack carrier as topology-constrained B\'{e}zier Curved-Blocks, providing a compact and manufacturable geometric parameterization, and optimizes one shared perturbation with Particle Swarm Optimization (PSO) under Thin Plate Spline (TPS) deformation and Expectation over Transformation (EOT) imaging transformations. The optimized pattern is then deployed as wearable cold patches without sample-specific re-optimization during deployment. Experiments on five infrared datasets and nine pedestrian detectors show consistent digital attack performance, strong cross-dataset and cross-model transferability, and a 92.59\% attack success rate in real-world physical experiments. Ablation, visualization, and defense analyses show that Curved-Blocks disrupt pedestrian thermal feature aggregation and remain difficult for image-restoration-style defenses to remove. These results reveal a practical universal physical vulnerability in current infrared pedestrian detection systems and provide a benchmark for robustness evaluation.
arXiv:2603.22169v3 Announce Type: replace
Abstract: We introduce VersualRL, a closed-loop framework for task-level robot planning that uses visual execution feedback to iteratively refine executable Behavior Trees through structured natural-language critique. VersualRL combines a vision-language model critic, which analyzes visual observations and Behavior Tree execution traces, with a large language model actor that performs discrete and interpretable policy updates. During physical deployment, both foundation models remain frozen, while the executed Behavior Tree is updated at the symbolic level without online gradient-based policy optimization. This enables transparent reasoning, explicit causal feedback, and human-interpretable policy evolution. We validate VersualRL on a real mobile robot performing a multi-stage navigation and manipulation task under execution uncertainty. Experimental results show that the framework supports explainable policy improvements, closed-loop adaptation to execution failures, and successful deployment on physical robotic systems.
arXiv:2603.23101v3 Announce Type: replace
Abstract: Intelligent spectroscopy serves as a pivotal element in AI-driven closed-loop scientific discovery, functioning as the critical bridge between matter structure and artificial intelligence. However, conventional expert-dependent spectral interpretation encounters substantial hurdles, including susceptibility to human bias and error, dependence on limited specialized expertise, and variability across interpreters. To address these challenges, we propose SpecXMaster, an intelligent framework leveraging Agentic Reinforcement Learning (RL) for NMR molecular spectral interpretation. SpecXMaster enables automated extraction of multiplicity information from both 1H and 13C spectra directly from raw FID (free induction decay) data. This end-to-end pipeline enables fully automated interpretation of NMR spectra into chemical structures. It demonstrates superior performance across multiple public NMR interpretation benchmarks and has been refined through iterative evaluations by professional chemical spectroscopists. We believe that SpecXMaster, as a novel methodological paradigm for spectral interpretation, will have a profound impact on the organic chemistry community.
arXiv:2507.18780v2 Announce Type: replace
Abstract: We consider data-driven reduced-order models of partial differential equations with shift equivariance. Shift-equivariant systems typically admit traveling solutions, and the main idea of our approach is to represent the solution in a traveling reference frame, in which it can be described by a relatively small number of basis functions. Existing methods for operator inference allow one to approximate a reduced-order model directly from data, without knowledge of the full-order dynamics. Our method adds additional terms to ensure that the reduced-order model not only approximates the spatially frozen profile of the solution, but also estimates the traveling speed as a function of that profile. We validate our approach using the Kuramoto-Sivashinsky equation, a one-dimensional partial differential equation that exhibits traveling solutions and spatiotemporal chaos. Results indicate that our method robustly captures traveling solutions, and exhibits improved numerical stability over the standard operator inference approach.
arXiv:2511.20956v2 Announce Type: replace
Abstract: Breast ultrasound (BUS) reporting relies on clinically meaningful lesion descriptors, including BI-RADS category, lesion shape, margin, echogenicity, posterior features, pathology, and histology. However, many public BUS datasets provide structured annotations and lesion masks without paired radiologist-written reports, limiting the development of vision--language models for BUS report generation. We propose BUSTR, a descriptor-aware vision--language framework that uses structured lesion information to enable report generation under limited report supervision. BUSTR first constructs descriptor-derived reports from available annotations and radiomics features extracted from lesion masks. It then trains a multi-head Swin Transformer encoder with multitask supervision to learn descriptor-aware visual representations across datasets with partially overlapping annotation sets. The projected visual tokens condition a frozen LLaMA-based language model, and training is guided by a dual-level objective combining token-level cross-entropy with representation-level cosine alignment. At inference, BUSTR generates reports from BUS images without access to structured descriptors, lesion masks, or radiomics features. We evaluate BUSTR on the public BrEaST and BUS-BRA datasets using natural language generation and clinical efficacy metrics. BUSTR improves report similarity and descriptor recovery compared with representative report-generation baselines, with notable gains for lesion shape, margin, posterior features, and pathology, as well as improved BI-RADS sensitivity and F1-score on BrEaST. These results suggest that structured BUS descriptors, lesion masks, and radiomics features can provide useful supervision for descriptor-aware BUS report generation when paired radiologist-written reports are unavailable.
arXiv:2601.08662v3 Announce Type: replace
Abstract: This tutorial is designed to make reinforcement learning (RL) more accessible to undergraduate students by offering clear, example-driven explanations. It focuses on bridging the gap between RL theory and practical coding applications, addressing common challenges that students face when transitioning from conceptual understanding to implementation. Through hands-on examples and approachable explanations, the tutorial aims to equip students with the foundational skills needed to confidently apply RL techniques in real-world scenarios.
arXiv:2603.24002v3 Announce Type: replace
Abstract: Physics-Informed Neural Networks (PINNs) for high-dimensional and high-order partial differential equations (PDEs) are primarily constrained by the $\mathcal{O}(d^k)$ spatial derivative complexity and the $\mathcal{O}(P)$ memory overhead of backpropagation (BP). While randomized spatial estimators successfully reduce the spatial complexity to $\mathcal{O}(1)$, their reliance on first-order optimization still leads to prohibitive memory consumption at scale. Zeroth-order (ZO) optimization offers a BP-free alternative; however, naively combining randomized spatial operators with ZO perturbations triggers a variance explosion of $\mathcal{O}(1/\varepsilon^2)$, leading to numerical divergence. To address these challenges, we propose the \textbf{S}tochastic \textbf{D}imension-free \textbf{Z}eroth-order \textbf{E}stimator (\textbf{SDZE}), a unified framework that achieves dimension-independent complexity in both space and memory. Specifically, SDZE leverages \emph{Common Random Numbers Synchronization (CRNS)} to algebraically cancel the $\mathcal{O}(1/\varepsilon^2)$ variance by locking spatial random seeds across perturbations. Furthermore, an \emph{implicit matrix-free subspace projection} is introduced to reduce parameter exploration variance from $\mathcal{O}(P)$ to $\mathcal{O}(r)$ while maintaining an $\mathcal{O}(1)$ optimizer memory footprint. Empirical results demonstrate that SDZE enables the training of 10-million-dimensional PINNs on a single NVIDIA A100 GPU, delivering significant improvements in speed and memory efficiency over state-of-the-art baselines.
arXiv:2603.24357v2 Announce Type: replace
Abstract: Robotic manipulation in unstructured environments requires end-effectors that combine kinematic dexterity with physical compliance. While traditional rigid hands rely on complex external sensors for safe interaction, electrohydraulic actuators offer a promising alternative by combining muscle-like compliance with self-sensing capability. This paper presents the design, control, and evaluation of a musculoskeletal robotic hand architecture powered entirely by remote Peano-HASEL actuators, optimized for safe manipulation. By relocating the actuators to the forearm, we isolate the grasping interface from electrical hazards while maintaining a slim, human-like profile. To address the inherently limited linear contraction of these soft actuators, we integrate a 1:2 pulley routing mechanism that mechanically amplifies tendon displacement. The resulting system prioritizes compliant interaction over high payload capacity, leveraging the intrinsic force-limiting characteristics of the actuators to provide inherent safety. This physical safety is augmented by the self-sensing nature of the HASEL actuators: by monitoring the operating current alone, we achieve real-time grasp detection and closed-loop contact-aware control without external force transducers or encoders. Experimental results demonstrate the system's dexterity and safety through the execution of grasp types from standard taxonomies and the non-destructive grasping of highly fragile objects such as a paper balloon. These findings represent a step toward simplified, inherently compliant soft robotic manipulation.
arXiv:2603.25468v2 Announce Type: replace
Abstract: This paper investigates how eight disciplinary research data repositories from the geosciences and social sciences navigate metadata conflicts - conflicts in implementations of the same standard and inter-standard conflicts - and how these conflicts affect the completeness of DataCite metadata. It combines results from analyzing DataCite metadata records, structural differences between three disciplinary metadata schemas and the DataCite Metadata Schema, and a direct comparison of two metadata records describing the same dataset. The results show that both conflicts in implementations of the same standard and inter-standard conflicts contribute to incomplete DataCite metadata. In addition to inherent differences between the metadata schemas, workflows and conscious decisions by the repositories also contribute to these conflicts. Some of these conflicts could be resolved by updated metadata crosswalks, emerging initiatives for retroactive collaborative metadata enrichment, the implementation of metadata application profiles or schema updates. Other conflicts can't be resolved or are strongly connected to the repository mission. The results also highlight that metadata completeness is multifaceted, and assessment requires careful consideration of context.
arXiv:2603.04233v2 Announce Type: replace
Abstract: Conventional cardiovascular magnetic resonance (CMR) cine imaging combines data across multiple heartbeats, an assumption that can fail in arrhythmia because beat-to-beat variation causes motion artifacts and obscures functional heterogeneity. Although 2D real-time cine resolves individual beats, stacked slices are suboptimal for capturing complex 3D cardiac dynamics. We investigated continuous beat-to-beat volumetric quantification in patients with premature ventricular contractions (PVCs) using free-running 3D real-time CMR. CMR-MOTUS was extended to jointly reconstruct time-resolved 3D motion fields and a motion-corrected reference image from continuously acquired data without breath-holding or ECG gating. Data were acquired with a variable-density Cartesian trajectory and either 3D spoiled gradient-echo or balanced steady-state free-precession imaging. Ventricular volumes were obtained by propagating one manual reference-image segmentation through all reconstructed frames. The method was evaluated in a cardiac motion phantom with static ground-truth acquisitions, 10 healthy volunteers, and 10 patients with PVCs; all in vivo scans were performed after contrast administration. Phantom ejection fraction (EF) agreed closely with ground truth (17.86% versus 17.27%). Healthy volunteers showed narrow beat-to-beat EF distributions, whereas patients with PVCs showed broader and sometimes bimodal distributions. Simultaneous ECG recordings supported the temporal correspondence between volume irregularities and PVC episodes. Free-running joint 3D motion-field and image reconstruction enables continuous beat-to-beat volumetric assessment and can reveal functional heterogeneity obscured by gated or heartbeat-averaged methods. Larger studies are required to establish clinical validity and determine its role alongside standard 2D cine analysis.
arXiv:2603.25623v3 Announce Type: replace
Abstract: Robust scene representation is essential for autonomous systems to safely operate in challenging low-visibility environments. In these conditions, radar has a clear advantage over cameras and lidars due to its resilience to environmental factors such as fog, smoke, or dust. However, radar data is inherently sparse and noisy, making reliable 3D surface reconstruction challenging. To address this, we propose a neural implicit approach for 3D mapping from radar point clouds that jointly models scene geometry and view-dependent radar intensities. Our method leverages a memory-efficient hybrid feature encoding to learn a continuous Signed Distance Field (SDF) for surface reconstruction, while also capturing radar-specific reflective properties. We show that our approach produces smoother, more accurate 3D surface reconstructions compared to existing lidar-based reconstruction methods applied to radar data and can reconstruct view-dependent radar intensities. We also show that, in general, as input point clouds get sparser, neural implicit representations render more faithful surfaces than traditional explicit SDFs and meshing techniques.
arXiv:2601.03043v4 Announce Type: replace
Abstract: Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term ``Less is Less'' (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks.
arXiv:2601.09761v3 Announce Type: replace
Abstract: It is shown that equations describing the Galilean electromagnetism in the presence of sources hold invariant under the l-conformal Galilei group for an arbitrary (half)integer parameter l. The group contains transformations which link an inertial frame of reference to those moving with constant accelerations of order up to 2l-1, thus pointing at potential dynamical instability.
arXiv:2603.07530v2 Announce Type: replace
Abstract: In-context imitation learning enables robots to adapt to new tasks from a small number of demonstrations without additional training. However, existing approaches typically condition only on state-action trajectories and lack explicit representations of task intent. This limitation hinders performance in complex and ambiguous task settings where the same actions may be consistent with different objectives. To address this, we present In-Context Imitation Learning with Visual Reasoning (ICLR), a novel framework that augments demonstration prompts with structured visual reasoning traces representing anticipated future robot trajectories in image space. ICLR also jointly learns to generate reasoning traces and low-level actions within a unified autoregressive transformer, enabling the model to mimic not only action prediction but also the reasoning process that leads to those actions. We extensively evaluate ICLR in both simulation and real-world manipulation tasks and demonstrate consistent improvements in success rates and generalization to unseen tasks and novel object configurations compared to other in-context imitation learning methods. These results suggest that incorporating embodied visual reasoning represents a promising direction for enhancing the robustness and generalization of robotic in-context learning systems.
arXiv:2603.25828v2 Announce Type: replace
Abstract: Hexagonally-packed polystyrene microsphere monolayers coated with tantalum pentoxide (Ta$_2$O$_5$) form scalable dielectric metasurfaces that support tunable photonic resonances and enhanced local density of optical states (LDOS). Here we combine fabrication, optical and fluorescence spectroscopy, and multiscale electromagnetic simulations to quantify how the thickness of the Ta$_2$O$_5$ shells control far-field resonances and Rhodamine 6G (Rh6G) emission. Experimentally, Ta$_2$O$_5$ shells of 10 - 70 nm deposited on microsphere lattices generate resonances that shift red with the thickness of the shell and systematically enhance the Rh6G fluorescence relative to flat Ta$_2$O$_5$ films. The largest enhancement is obtained for 30 - 50 nm shells, when lattice resonances overlap the Rh6G excitation and emission bands. Finite-cluster finite-difference time-domain simulations reproduce the measured transmittance and reflectance spectra, confirming the assumed geometry of the Ta$_2$O$_5$ shells covering the sphere lattice. Periodic-cell simulations of single electric dipoles yield wavelength-dependent Purcell factors $Fp(\lambda)$ and directional $\beta$-factors $\beta_{top}(\lambda)$, from which we construct emission-weighted figures of merit that link LDOS modulation to the experimentally accessible top-side fluorescence enhancement. As a complementary test of our emitter-environment model, we compare simulated and measured Purcell factors for PS/Ta$_2$O$_5$ microsphere lattices. A physically motivated averaging that accounts for emitter position, orientation and ensemble spectral smoothing yields very good agreement across all shells. Overall, our results establish Ta$_2$O$_5$-coated microsphere lattices as robust dielectric substrates for surface-enhanced fluorescence and clarify how shell thickness and emitter placement jointly control photonic resonances, LDOS and fluorescence response.
arXiv:2503.17657v3 Announce Type: replace
Abstract: Pretrained diffusion models exhibit a well-known training-sampling mismatch, often attributed to exposure bias and related distribution-shift effects. We provide a quantitative interpretation of this phenomenon through the notion of an effective noise level: empirically, a pretrained denoiser behaves as if trained at a noise level slightly below the nominal schedule. Motivated by this observation, we introduce a training-free sampling strategy based on truncating the Karhunen--Lo\`eve (KL) expansion of the Brownian motion driving the forward stochastic differential equation. Truncation yields a finite-dimensional forward process with a reduced noise level that can be adjusted independently of the time discretization. We prove uniform convergence of the truncated process to the original diffusion. To explain the resulting behaviour, we analyse a toy model in which the denoiser is exact but operates at a reduced effective noise level. The analysis predicts a non-monotone response to the sampling noise with a unique interior optimum, located by a one-dimensional sweep over the truncation order. We implement the approach through corresponding truncated reverse-time and probability-flow equations, without modifying the network architecture. Across CIFAR-10, CelebA, ImageNet, and latent-space Stable Diffusion, the truncation order consistently reveals a sweet spot, improving pretrained models in nearly all tested configurations. Training from scratch at a matched truncation order makes that order the network's own sweet spot, accelerates convergence by about $2.8\times$, and lowers the generation error on CIFAR-10 (Fr\'echet Inception Distance 7.14 to 5.47 at matched epochs; best checkpoint 6.74 to 5.23). A L\'evy--Ciesielski comparison confirms that finite expansion is broadly beneficial.
arXiv:2603.26648v3 Announce Type: replace
Abstract: Recent advances in large language models have improved the capabilities of coding agents, yet systematic evaluation of complex, end-to-end website development remains limited. To address this gap, we introduce Vision2Web, a hierarchical benchmark for visual website development, spanning from static UI-to-code generation, interactive multi-page frontend reproduction, to long-horizon full-stack website development. The benchmark is constructed from real-world websites and comprises a total of 193 tasks across 16 categories, with 918 prototype images and 1,255 test cases. To support flexible, thorough and reliable evaluation, we propose workflow-based agent verification paradigm based on two complementary components: a GUI agent verifier and a VLM-based judge. We evaluate multiple visual language models instantiated under different coding-agent frameworks, revealing substantial performance gaps at all task levels, with state-of-the-art models still struggling on full-stack development.