Searcharxiv⌕ Search

arXiv subjects

Di Liu

Publications and source records attributed to Di Liu.

At least 37 records · Page 2Linked to original sources
arXiv2026

EdgeCompress: Coupling Multidimensional Model Compression and Dynamic Inference for EdgeAI

Convolutional neural networks (CNNs) have demonstrated encouraging results in image classification tasks. However, the prohibitive computational cost of CNNs hinders the deployment of CNNs onto resource-constrained embedded devices. To address this issue, we propose EdgeCompress, a comprehensive compression framework to reduce the computational overhead of CNNs. In EdgeCompress, we first introduce dynamic image cropping (DIC), where we design a lightweight foreground predictor to accurately crop the most informative foreground object of input images for inference, which avoids redundant computation on background regions. Subsequently, we present compound shrinking (CS) to collaboratively compress the three dimensions (depth, width, and resolution) of CNNs according to their contribution to accuracy and model computation. DIC and CS together constitute a multidimensional CNN compression framework, which is able to comprehensively reduce the computational redundancy in both input images and neural network architectures, thereby improving the inference efficiency of CNNs. Further, we present a dynamic inference framework to efficiently process input images with different recognition difficulties, where we cascade multiple models with different complexities from our compression framework and dynamically adopt different models for different input images, which further compresses the computational redundancy and improves the inference efficiency of CNNs, facilitating the deployment of advanced CNNs onto embedded hardware. Experiments on ImageNet-1K demonstrate that EdgeCompress reduces the computation of ResNet-50 by 48.8% while improving the top-1 accuracy by 0.8%. Meanwhile, we improve the accuracy by 4.1% with similar computation compared to HRank, the state-of-the-art compression framework. The source code and models are available at https://github.com/ntuliuteam/edge-compress

cs.CV↗
arXiv2026

Smart Scissor: Coupling Spatial Redundancy Reduction and CNN Compression for Embedded Hardware

Scaling down the resolution of input images can greatly reduce the computational overhead of convolutional neural networks (CNNs), which is promising for edge AI. However, as an image usually contains much spatial redundancy, e.g., background pixels, directly shrinking the whole image will lose important features of the foreground object and lead to severe accuracy degradation. In this paper, we propose a dynamic image cropping framework to reduce the spatial redundancy by accurately cropping the foreground object from images. To achieve the instance-aware fine cropping, we introduce a lightweight foreground predictor to efficiently localize and crop the foreground of an image. The finely cropped images can be correctly recognized even at a small resolution. Meanwhile, computational redundancy also exists in CNN architectures. To pursue higher execution efficiency on resource-constrained embedded devices, we also propose a compound shrinking strategy to coordinately compress the three dimensions (depth, width, resolution) of CNNs. Eventually, we seamlessly combine the proposed dynamic image cropping and compound shrinking into a unified compression framework, Smart Scissor, which is expected to significantly reduce the computational overhead of CNNs while still maintaining high accuracy. Experiments on ImageNet-1K demonstrate that our method reduces the computational cost of ResNet50 by 41.5% while improving the top-1 accuracy by 0.3%. Moreover, compared to HRank, the state-of-the-art CNN compression framework, our method achieves 4.1% higher top-1 accuracy at the same computational cost. The codes and data are available at https://github.com/ntuliuteam/smart-scissor

cs.CV↗
arXiv2026

SceneFrom3D: Geometry-Conditioned Outdoor 3D Scene Generation via View Scheduling with Object-Level Control

Geometry-conditioned 3D scene generation enables the creation of 3D environments from user-provided geometry, offering direct control over scene structure and object layout. To generate such 3D scenes, current methods commonly adopt a three-stage design that first defines a view schedule, then synthesizes multi-view observations along the scheduled views, and finally reconstructs a 3D representation from the generated images. However, defining the view schedule becomes a major bottleneck for outdoor scenes, where large, unstructured, and unbounded geometry makes it difficult to obtain views that provide sufficient coverage while supporting stable generation. To address this bottleneck, we present SceneFrom3D, a framework that automatically schedules views from outdoor input geometries. SceneFrom3D constructs a directed generation graph whose nodes represent anchor views and whose edges represent interpolation trajectories, defining which views to synthesize, which view pairs to interpolate, and in which order generation should proceed. Beyond automatic view scheduling, SceneFrom3D further improves controllability through object-level conditioning, assigning each object an identity image for appearance guidance and a geometry-adherence parameter for region-wise control over the input geometry. Experiments demonstrate that SceneFrom3D achieves state-of-the-art geometry-conditioned outdoor 3D scene generation, producing high-quality scenes with controllable object appearance and geometry adherence.

cs.GR↗
arXiv2026

Linear Power System Modeling and Analysis Across Wide Operating Ranges: A Hierarchical Neural State-Space Equation Approach

As modern power systems exhibit increasingly high-dimensional, nonlinear, and uncertain characteristics, the applicability of classical linear state-space methods is severely challenged. Existing paradigms struggle to reconcile the analytical transparency of physics-based models with the continuous nonlinear generalization of AI. To address this, the Hierarchical Neural State-Space Equation (HNSSE) framework is proposed. At the component level, the formulated Neural State-Space Equation (NSSE) extends neural ordinary differential equations (NODEs) to learn continuous dynamic manifolds across varying conditions while strictly preserving local analytical transparency. At the system level, a hierarchical architecture analytically fuses components via network constraints, constructing an interaction-consistent global NODE while circumventing the curse of dimensionality. To ensure robust convergence under noisy measurements, a training strategy synergizing spatiotemporal slicing, physics-informed curriculum learning, and Expectation-Maximization-based refinement is established. Validation on the large-scale Guangdong Power Grid demonstrates the framework's remarkable performance in interpretable state-space reconstruction, high-fidelity trajectory prediction, continuous stability perception, and noise robustness. Comprehensive comparisons substantiate HNSSE's superiority as a unified, interpretable paradigm for complex power system modeling.

eess.SY↗
arXiv2026

Multigrid Training for Molecular Generation using Graph Neural Networks

Deep learning has demonstrated significant success for modeling biochemical molecular systems, where inputs are commonly represented as graphs or 3D grids. A major challenge is that computational cost scales with resolution, making full graph/grid computation of molecular densities expensive and often unstable. We introduce a multigrid training strategy that leverages low-resolution optimization to accelerate learning at higher resolution through parameter transfer across discretizations. For graph molecular representations, we progressively transfer parameters learned from a coarse graph to a sequence of increasingly finer graphs via biased random walk upsampling. For 3D molecular generation, we voxelize the molecular structures at multiple resolutions, pretrain a coarse-resolution conditional Variational Autoencoder (CVAE), and initialize a fine-resolution CVAE by transferring shape compatible convolutional parameters from the coarse model. Numerical experiments on receptor-conditioned 3D Ligand generation show that multigrid training accelerates convergence and improves generalization compared to training from scratch.

cs.LG↗
arXiv2026

AB-Sparse: Sparse Attention with Adaptive Block Size for Accurate and Efficient Long-Context Inference

As large language models scale to longer contexts, loading the growing KV cache during attention computation becomes a critical bottleneck. Previous work has shown that attention computation is dominated by a small subset of tokens. This motivates block sparse attention methods that partition the KV cache into fixed-size blocks and selectively compute attention over those blocks exhibiting high importance. However, these methods assign a uniform block size across all attention heads, implicitly assuming homogeneous behavior throughout the model. Our analysis reveals that this assumption is flawed: attention heads exhibit widely varying sensitivity to block granularity, and uniformity leads to suboptimal accuracy. We present AB-Sparse, a training-free algorithm-system co-designed framework that improves accuracy while preserving throughput. AB-Sparse introduces lightweight adaptive block size allocation across attention heads to improve accuracy. To compensate for the additional memory overhead, it further employs lossless block centroid quantization. In addition, custom GPU kernels are developed to support efficient execution with variable block sizes. Evaluation results demonstrate that AB-Sparse achieves an accuracy improvement of up to 5.43% over existing block sparse attention baselines without throughput overhead.

cs.DC↗
arXiv2026

RetroInfer: A Vector Storage Engine for Scalable Long-Context LLM Inference

Recent large language models (LLMs) are rapidly extending their context windows, yet inference throughput lags due to increasing GPU memory and bandwidth demands. This is because the key-value (KV) cache, an intermediate structure storing token representations, grows linearly with context length and requires an iterative linear scan for attention computation. A promising direction to accelerate long-context inference is to exploit attention's inherent sparsity by offloading the KV cache to CPU memory and retrieving only a small subset of tokens important to the current generation step. However, prior sparse attention approaches struggle to balance accuracy and retrieval cost due to varying sparsity patterns and inefficient GPU-CPU memory management. We present RetroInfer, a vector storage engine that realizes a sparsity-based KV cache for long-context inference. RetroInfer introduces an Attention-aWare VEctor index (wave index), which fundamentally improves the tradeoff between attention accuracy and retrieval cost through tripartite attention approximation, accuracy-bound attention estimation, and segmented clustering. We also design the wave buffer, a GPU-CPU buffer manager that assigns computation and manages data across heterogeneous hardware. We evaluate RetroInfer across a range of models and workloads, demonstrating up to 4.4X decoding throughput over full attention at 120K context and up to 12.2X over sparse attention baselines at 1 million tokens -- all while preserving full-attention-level accuracy.

cs.LG↗
arXiv2026

Extending Linear Convergence of the Proximal Point Algorithm: The Quasar-Convex Case

This work investigates the properties of the proximity operator for quasar-convex functions and establishes the convergence of the proximal point algorithm to a global minimizer with a particular focus on its convergence rate. In particular, we demonstrate: (i) the generated sequence is mi\-ni\-mi\-zing and achieves an $\mathcal{O}(\varepsilon^{-1})$ complexity rate for quasar-convex functions; (ii) under strong quasar-convexity, the sequence converges linearly and attains an $\mathcal{O}(\ln(\varepsilon^{-1}))$ complexity rate. These results extend known convergence rates from the (strongly) convex to the (strongly) quasar-convex setting. To the best of our knowledge, some findings are novel even for the special case of (strongly) star-convex functions. Numerical experiments corroborate our theoretical results.

math.OC↗
arXiv2026

C2RustXW: Program-Structure-Aware C-to-Rust Translation via Program Analysis and LLM

The growing adoption of Rust for its memory safety and performance has increased the demand for effective migration of legacy C codebases. However, existing rule-based translators (e.g., \ctorust) often generate verbose, non-idiomatic code that preserves unsafe C semantics, limiting readability, maintainability, and practical adoption. Moreover, manual post-processing of such outputs is labor-intensive and rarely yields high-quality Rust code, posing a significant barrier to large-scale migration. To address these limitations, we present \tool, a program-structure-aware C-to-Rust translation approach that integrates program analysis with Large Language Models (LLMs). \tool extracts the multi-level program structure, including global symbols, function dependencies, and control- and data-flow information, and encodes these as structured textual representations injected into LLM prompts to guide translation and repair. Based on this design, \tool performs dependency-aware translation and adopts a multi-stage repair pipeline that combines rule-based and structure-guided LLM-based techniques to ensure syntactic correctness. For semantic correctness, \tool further integrates execution-based validation with structure-guided reasoning to localize and repair behavioral inconsistencies. Experimental results show that \tool achieves 100\% syntactic correctness on CodeNet and 97.78\% on GitHub, while significantly reducing code size (up to 43.70\%) and unsafe usage (to 5.75\%). At the project level, \tool achieves perfect syntactic correctness and an average semantic correctness of 78.87\%, demonstrating its effectiveness for practical and scalable C-to-Rust migration.

cs.SE↗
arXiv2026

S-HPLB: Efficient LLM Attention Serving via Sparsity-Aware Head Parallelism Load Balance

With the increasing volumes of Large Language Models (LLMs) and the expanding context lengths, attention computation has become a key performance bottleneck in LLM serving. For fast attention computation, recent practices often parallelize the attention heads on multiple GPUs, and also widely adopt attention sparsification to reduce the computation amount -- which selectively computes a subset of attention pairs under a preset sparsity budget. In this paper, we notice that attention heads of an LLM model often exhibit heterogeneous-yet-stable sparsity elasticities, which motivates us to enforce head-adaptive sparsity budgets to attain better efficiency while preserving high inference quality. Yet, from the system aspect, with heterogeneous sparsity levels, attention computation time on different heads would be inconsistent, yielding cross-GPU resource bubbles under head-parallel deployment. To further minimize such bubbles, we propose a novel attention deployment strategy called Sparsity-aware Head-Parallel Load Balance (S-HPLB). Experiments on long-context benchmark show that, S-HPLB can achieve a $2.88\times$ improvement in average attention computation latency without quality degradation.

cs.DC↗
arXiv2026

Snapmoji: Instant Generation of Animatable Dual-Stylized Avatars

Despite the increasing popularity of avatar systems such as Snapchat Bitmojis, existing production avatar platforms face several limitations, such as a limited number of predefined assets, tedious customization processes, and inefficient rendering requirements. Addressing these shortcomings, we introduce Snapmoji, an avatar generation system that instantly creates 3D avatars, and enables customization in a process we call dual-stylization. Snapmoji first maps a selfie of a user to a primary avatar (e.g., Bitmoji style) using a new technique we name Gaussian Domain Adaptation (GDA), then applies a secondary style (e.g., skeleton, yarn, toy) to the primary avatar, all while preserving the user's identity. The generated 3D avatars can then be rendered an animated on mobile devices at 30-40 FPS.

cs.GR↗
arXiv2026

On circumcentered direct methods for monotone variational inequality problems

Circumcentered techniques have been shown to significantly accelerate projection-based methods for convex feasibility problems. Motivated by this success, we propose two direct methods with circumcenter acceleration for solving variational inequality problems involving two classes of operators: paramonotone and monotone. Both schemes rely on approximate projections onto separating halfspaces, thereby avoiding computationally expensive exact projections. We establish convergence results for both methods and conduct numerical experiments, demonstrating that the proposed algorithms outperform classical methods, such as the extragradient algorithm, by orders of magnitude in terms of computational time, particularly when the feasible set is a complex intersection of convex sets.

math.OC↗
arXiv2026

When Generative AI Is Intimate, Sexy, and Violent: Examining Not-Safe-For-Work (NSFW) Chatbots on FlowGPT

User-created chatbots powered by generative AI offer new ways to share and interact with Not-Safe-For-Work (NSFW) content. However, little is known about the characteristics of these GenAI-based chatbots and their user interactions. Drawing on the functional theory of NSFW on social media, this study analyzes 376 NSFW chatbots and 307 public conversation sessions on FlowGPT. Findings identify four chatbot types: roleplay characters, story generators, image generators, and do-anything-now bots. AI Characters portraying fantasy personas and enabling hangout-style interactions are most common, often using explicit avatar images to invite engagement. Sexual, violent, and insulting content appears in both user prompts and chatbot outputs, with some chatbots generating explicit material even when users do not create erotic prompts. In sum, the NSFW experience on FlowGPT can be understood as a combination of virtual intimacy, sexual delusion, violent thought expression, and unsafe content acquisition. We conclude with implications for chatbot design, creator support, user safety, and content moderation.

cs.HC↗
arXiv2025

Intrinsic Meron Spin Textures in Generic Focused Fields

Optical spin textures with nontrivial topology hold promise for structured light and photonic information processing, yet their generation typically relies heavily on externally structured light with care. This raises questions about their universal existence and true robustness. Here, we uncover and experimentally verify a meron-like spin texture that emerges intrinsically in focused fields, without any wavefront engineering. This intrinsic meron spin texture, unlike their externally engineered counterparts, exhibits exceptional robustness against a wide range of inputs, including partially polarized and spatially disordered pupils corrupted by decoherence and depolarization. We attribute its resilience to topological protection from phase vortices in the focal field. Our findings reveal a naturally occurring spin structure that is intrinsic to the focused field with exceptional robustness against noise, which complements the existing externally engineered ones. It offers new ingredients into topological spin textures in optics and enriches their potentials for disorder-resilient photonic applications.

physics.optics↗
arXiv2025

Efficient Deep Learning Infrastructures for Embedded Computing Systems: A Comprehensive Survey and Future Envision

Deep neural networks (DNNs) have recently achieved impressive success across a wide range of real-world vision and language processing tasks, spanning from image classification to many other downstream vision tasks, such as object detection, tracking, and segmentation. However, previous well-established DNNs, despite being able to maintain superior accuracy, have also been evolving to be deeper and wider and thus inevitably necessitate prohibitive computational resources for both training and inference. This trend further enlarges the computational gap between computation-intensive DNNs and resource-constrained embedded computing systems, making it challenging to deploy powerful DNNs upon real-world embedded computing systems towards ubiquitous embedded intelligence. To alleviate the above computational gap and enable ubiquitous embedded intelligence, we, in this survey, focus on discussing recent efficient deep learning infrastructures for embedded computing systems, spanning from training to inference, from manual to automated, from convolutional neural networks to transformers, from transformers to vision transformers, from vision models to large language models, from software to hardware, and from algorithms to applications. Specifically, we discuss recent efficient deep learning infrastructures for embedded computing systems from the lens of (1) efficient manual network design for embedded computing systems, (2) efficient automated network design for embedded computing systems, (3) efficient network compression for embedded computing systems, (4) efficient on-device learning for embedded computing systems, (5) efficient large language models for embedded computing systems, (6) efficient deep learning software and hardware for embedded computing systems, and (7) efficient intelligent applications for embedded computing systems.

cs.LG↗
arXiv2025

Detecting Ultralight Dark Matter with Matter Effect

Ultralight particles, with a mass below the electronvolt scale, exhibit wave-like behavior and have arisen as a compelling dark matter candidate. A particularly intriguing subclass is scalar dark matter, which induces variations in fundamental physical constants. However, detecting such particles becomes highly challenging in the mass range above $10^{-6}\,\text{eV}$, as traditional experiments face severe limitations in response time. In contrast, the matter effect becomes significant in a vast and unexplored parameter space. These effects include (i) a force arising from scattering between ordinary matter and the dark matter wind and (ii) a fifth force between ordinary matter induced by the dark matter background. Using the repulsive quadratic scalar-photon interaction as a case study, we develop a unified framework based on quantum mechanical scattering theory to systematically investigate these phenomena across both perturbative and non-perturbative regimes. Our approach not only reproduces prior results obtained through other methodologies but also covers novel regimes with nontrivial features, such as decoherence effects, screening effects, and their combinations. In particular, we highlight one finding related to both scattering and background-induced forces: the descreening effect observed in the non-perturbative region with large incident momentum, which alleviates the decoherence suppression. Furthermore, we discuss current and proposed experiments, including inverse-square-law tests, equivalence principle tests, and deep-space acceleration measurements. Notably, we go beyond the spherical approximation and revisit the MICROSCOPE constraint on the background-induced force in the large-momentum regime, where the decoherence and screening effects interplay. The ultraviolet models realizing the quadratic scalar-photon interaction are also discussed.

hep-ph↗
arXiv2025

Modulation of Ionic Current Rectification in Short Unipolar Nanopores

With controlled ionic current rectification (ICR) achieved through a strategically designed non-uniform surface charge distribution, short unipolar nanopores exhibit promising applications in nanofluidic sensors, ionic circuits, and ion amplifiers. By systematically investigating how the charged length on inner pore walls modulates ion transport, we found that both the maximum ICR degree and the corresponding charged-length proportion were influenced by nanopore parameters and simulation conditions. For 100 nm-long unipolar nanopores, the highest ICR degree is obtained at a charged-length proportion of ~0.3, due to the corresponding most significant ion enrichment and depletion inside the nanopore under opposite biases. This charged-length proportion of ~0.3 consistently appears as a characteristic value across most considered cases. For short unipolar nanopores, the presence of exterior surface charges significantly enhances the ICR degree by facilitating ion transport through nanopores. The effective widths of charged regions beyond nanopore borders on outer surfaces exhibit direct proportionality to the pore diameter, surface charge density, and applied voltage, and inverse proportionality to the pore length and salt concentration. Our research may provide useful guidance for the design of unipolar nanopores and porous membranes incorporating such charge configurations.

physics.chem-ph↗
arXiv2025

The Alignment Paradox of Medical Large Language Models in Infertility Care: Decoupling Algorithmic Improvement from Clinical Decision-making Quality

Large language models (LLMs) are increasingly adopted in clinical decision support, yet aligning them with the multifaceted reasoning pathways of real-world medicine remains a major challenge. Using more than 8,000 infertility treatment records, we systematically evaluate four alignment strategies: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), Group Relative Policy Optimization (GRPO), and In-Context Learning (ICL) through a dual-layer framework combining automatic benchmarks with blinded doctor-in-the-loop assessments. GRPO achieves the highest algorithmic accuracy across multiple decision layers, confirming the value of reinforcement-based optimization for structured prediction tasks. However, clinicians consistently prefer the SFT model, citing clearer reasoning processes (p = 0.035) and higher therapeutic feasibility (p = 0.019). In blinded pairwise comparisons, SFT attains the highest winning rate (51.2%), outperforming both GRPO (26.2%) and even physicians' original decisions (22.7%). These results reveal an alignment paradox: algorithmic improvements do not necessarily translate into higher clinical trust, and may diverge from human-centered preferences. Our findings highlight the need for alignment strategies that prioritize clinically interpretable and practically feasible reasoning, rather than solely optimizing decision-level accuracy.

cs.LG↗