Searcharxiv⌕ Search

arXiv subjects

Sheng Wang

Publications and source records attributed to Sheng Wang.

At least 55 records · Page 3Linked to original sources
arXiv2026

Relay Buffer Independent Communication over Pooled HBM for Efficient MoE Inference on Ascend

Mixture-of-Experts (MoE) inference requires large-scale token exchange across devices, making dispatch and combine major bottlenecks in both prefill and decode. Beyond network transfer, routing-driven layout transformation, temporary relay, and output restoration can add substantial overhead. Existing MoE communication paths are often buffer-centric, using explicit inter-process relay and reordering buffers around collective transfer. This report presents a relay-buffer-free communication design for MoE inference acceleration on Ascend systems. The design reorganizes dispatch and combine around direct placement into destination expert windows and direct reading from remote expert windows. Built on globally pooled high-bandwidth memory and symmetric-memory allocation, it removes most intermediate relay and reordering buffers while retaining only lightweight control state, including counts, offsets, and synchronization metadata. We instantiate the design as two schedules for the main phases of MoE inference: a prefill schedule with richer planning state for throughput-oriented execution, and a compact decode schedule for latency-sensitive execution. Experiments on Ascend-based MoE workloads show reduced dispatch and combine latency in both settings. At the serving level, the implementation improves time to first token (TTFT), preserves competitive time per output token (TPOT), and enlarges the feasible scheduling space under practical latency constraints. These results indicate that, on platforms with globally addressable device memory, reducing intermediate buffering and output restoration around expert execution is an effective direction for accelerating MoE inference.

cs.DC↗
arXiv2026

MicroDiffuse3D: A Foundation Model for 3D Microscopy Imaging Restoration

Chemical imaging enables label-free visualization of cells, tissues and living systems while providing direct biochemical information that is difficult to obtain with conventional fluorescence microscopy. Despite its promise in applications ranging from intraoperative diagnosis to drug-response analysis, its broader use remains limited by slow data acquisition, particularly for three-dimensional imaging. Here we present MicroDiffuse3D, a pretrained foundation model for 3D microscopy image restoration that recovers high-quality volumetric structure from degraded low-resolution measurements acquired at substantially higher throughput. We evaluated MicroDiffuse3D across three challenging restoration settings, including 3D super-resolution under 16-fold volumetric sparsity, joint degradation in resolution and noise, and 3D denoising in the low signal-to-noise ratio (SNR) regime, where the model delivered clear gains over strong baselines. Under the sparse 3D super-resolution setting, MicroDiffuse3D produced clearer continuity across depth with fewer artifacts and improved segmentation quality by 10.58% and line-profile concordance by 15.59%. Together, our results establish pretrained 3D restoration as a broadly applicable strategy for overcoming the throughput and SNR limitations in volumetric chemical imaging, enabling high-resolution analysis at scales and speeds that were previously difficult to achieve.

cs.CV↗
arXiv2026

AmodalSVG: Amodal Image Vectorization via Semantic Layer Peeling

We introduce AmodalSVG, a new framework for amodal image vectorization that produces semantically organized and geometrically complete SVG representations from natural images. Existing vectorization methods operate under a modal paradigm: tracing only visible pixels and disregarding occlusion. Consequently, the resulting SVGs are semantically entangled and geometrically incomplete, limiting SVG's structural editability. In contrast, AmodalSVG reconstructs full object geometries, including occluded regions, into independent, editable vector layers. To achieve this, AmodalSVG reformulates image vectorization as a two-stage framework, performing semantic decoupling and completion in the raster domain to produce amodally complete semantic layers, which are then independently vectorized. In the first stage, we introduce Semantic Layer Peeling (SLP), a VLM-guided strategy that progressively decomposes an image into semantically coherent layers. By hybrid inpainting, SLP recovers complete object appearances under occlusions, enabling explicit semantic decoupling. To vectorize these layers efficiently, we propose Adaptive Layered Vectorization (ALV), which dynamically modulates the primitive budget via an error-budget-driven adjustment mechanism. Extensive experiments demonstrate that AmodalSVG significantly outperforms prior methods in visual fidelity. Moreover, the resulting amodal layers enable object-level editing directly in the vector domain, capabilities not supported by existing vectorization approaches. Code will be released upon acceptance.

cs.CV↗
arXiv2026

SVGFusion: A VAE-Diffusion Transformer for Vector Graphic Generation

Generating high-quality Scalable Vector Graphics (SVGs) from text remains a significant challenge. Existing LLM-based models that generate SVG code as a flat token sequence struggle with poor structural understanding and error accumulation, while optimization-based methods are slow and yield uneditable outputs. To address these limitations, we introduce SVGFusion, a unified framework that adapts the VAE-diffusion architecture to bridge the dual code-visual nature of SVGs. Our model features two core components: a Vector-Pixel Fusion Variational Autoencoder (VP-VAE) that learns a perceptually rich latent space by jointly encoding SVG code and its rendered image, and a Vector Space Diffusion Transformer (VS-DiT) that achieves globally coherent compositions through iterative refinement. Furthermore, this architecture is enhanced by a Rendering Sequence Modeling strategy, which ensures accurate object layering and occlusion. Evaluated on our novel SVGX-Dataset comprising 240k human-designed SVGs, SVGFusion establishes a new state-of-the-art, generating high-quality, editable SVGs that are strictly semantically aligned with the input text.

cs.CV↗
arXiv2026

A Cross-graph Tuning-free GNN Prompting Framework

GNN prompting aims to adapt models across tasks and graphs without requiring extensive retraining. However, most existing graph prompt methods still require task-specific parameter updates and face the issue of generalizing across graphs, limiting their performance and undermining the core promise of prompting. In this work, we introduce a Cross-graph Tuning-free Prompting Framework (CTP), which supports both homogeneous and heterogeneous graphs, can be directly deployed to unseen graphs without further parameter tuning, and thus enables a plug-and-play GNN inference engine. Extensive experiments on few-shot prediction tasks show that, compared to SOTAs, CTP achieves an average accuracy gain of 30.8% and a maximum gain of 54%, confirming its effectiveness and offering a new perspective on graph prompt learning.

cs.LG↗
arXiv2026

Graph-centric Cross-model Data Integration and Analytics in a Unified Multi-model Database

Graph-centric cross-model data integration and analytics (GCDIA) refer to tasks that leverage the graph model as a central paradigm to integrate relevant information across heterogeneous data models, such as relational and document, and subsequently perform complex analytics such as regression and similarity computation. As modern applications generate increasingly diverse data and move beyond simple retrieval toward advanced analytical objectives (e.g., prediction and recommendation), GCDIA has become increasingly important. Existing multi-model databases (MMDBs) struggle to efficiently support both integration (GCDI) and analytics (GCDA) in GCDIA. They typically separate graph processing from other models without global optimization for GCDI, while relying on tuple-at-a-time execution for GCDA, leading to limited performance and scalability. To address these limitations, we propose GredoDB, a unified MMDB that natively supports storing graph, relational, and document models, while efficiently processing GCDIA. Specifically, we design 1) topology- and attribute-aware graph operators for efficient predicate-aware traversal, 2) a unified GCDI optimization framework to exploit cross-model correlations, and 3) a parallel GCDA architecture that materializes intermediate results for operator-level execution. Experiments on the widely adopted multi-model benchmark M2Bench demonstrate that, in terms of response time, GredoDB achieves up to 107.89 times and an average of 10.89 times speedup on GCDI, and up to 356.72 times and an average of 37.79 times on GCDA, compared to state-of-the-art (SOTA) MMDBs.

cs.DB↗
arXiv2026

Why Database Manuals Are Not Enough: Efficient and Reliable Configuration Tuning for DBMSs via Code-Driven LLM Agents

Modern database management systems (DBMSs) expose hundreds of configuration knobs that critically influence performance. Existing automated tuning methods either adopt a data-driven paradigm, which incurs substantial overhead, or rely on manual-driven heuristics extracted from database documentation, which are often limited and overly generic. Motivated by the fact that the control logic of configuration knobs is inherently encoded in the DBMS source code, we argue that promising tuning strategies can be mined directly from the code, uncovering fine-grained insights grounded in system internals. To this end, we propose SysInsight, a code-driven database tuning system that automatically extracts fine-grained tuning knowledge from DBMS source code to accelerate and stabilize the tuning process. SysInsight combines static code analysis with LLM-based reasoning to identify knob-controlled execution paths and extract semantic tuning insights. These insights are then transformed into quantitative and verifiable tuning rules via association rule mining grounded in tuning observations. During online tuning, system diagnosis is applied to identify critical knobs, which are adjusted under the rule guidance. Evaluations demonstrate that compared to the SOTA baseline, SysInsight converges to the best configuration on average 7.11X faster while achieving a 19.9% performance improvement.

cs.DB↗
arXiv2026

Let's Think with Images Efficiently! An Interleaved-Modal Chain-of-Thought Reasoning Framework with Dynamic and Precise Visual Thoughts

Recently, Interleaved-modal Chain-of-Thought (ICoT) reasoning has achieved remarkable success by leveraging both multimodal inputs and outputs, attracting increasing attention. While achieving promising performance, current ICoT methods still suffer from two major limitations: (1) Static Visual Thought Positioning, which statically inserts visual information at fixed steps, resulting in inefficient and inflexible reasoning; and (2) Broken Visual Thought Representation, which involves discontinuous and semantically incoherent visual tokens. To address these limitations, we introduce Interleaved-modal Chain-of-Thought reasoning with Dynamic and Precise Visual Thoughts (DaP-ICoT), which incorporates two key components: (1) Dynamic Visual Thought Integration adaptively introduces visual inputs based on reasoning needs, reducing redundancy and improving efficiency. (2) Precise Visual Thought Guidance ensures visual semantically coherent and contextually aligned representations. Experiments across multiple benchmarks and models demonstrate that DaP-ICoT achieves state-of-the-art performance. In addition, DaP-ICoT significantly reduces the number of inserted images, leading to a 72.6% decrease in token consumption, enabling more efficient ICoT reasoning.

cs.CV↗
arXiv2026

Optimizing Feature Extraction for On-device Model Inference with User Behavior Sequences

Machine learning models are widely integrated into modern mobile apps to analyze user behaviors and deliver personalized services. Ensuring low-latency on-device model execution is critical for maintaining high-quality user experiences. While prior research has primarily focused on accelerating model inference with given input features, we identify an overlooked bottleneck in real-world on-device model execution pipelines: extracting input features from raw application logs. In this work, we explore a new direction of feature extraction optimization by analyzing and eliminating redundant extraction operations across different model features and consecutive model inferences. We then introduce AutoFeature, an automated feature extraction engine designed to accelerate on-device feature extraction process without compromising model inference accuracy. AutoFeature comprises three core designs: (1) graph abstraction to formulate the extraction workflows of different input features as one directed acyclic graph, (2) graph optimization to identify and fuse redundant operation nodes across different features within the graph; (3) efficient caching to minimize operations on overlapping raw data between consecutive model inferences. We implement a system prototype of AutoFeature and integrate it into five industrial mobile services spanning search, video and e-commerce domains. Online evaluations show that AutoFeature reduces end-to-end on-device model execution latency by 1.33x-3.93x during daytime and 1.43x-4.53x at night.

cs.LG↗
arXiv2026

Supervised Fine-Tuning versus Reinforcement Learning: A Study of Post-Training Methods for Large Language Models

Pre-trained Large Language Model (LLM) exhibits broad capabilities, yet, for specific tasks or domains their attainment of higher accuracy and more reliable reasoning generally depends on post-training through Supervised Fine-Tuning (SFT) or Reinforcement Learning (RL). Although often treated as distinct methodologies, recent theoretical and empirical developments demonstrate that SFT and RL are closely connected. This study presents a comprehensive and unified perspective on LLM post-training with SFT and RL. We first provide an in-depth overview of both techniques, examining their objectives, algorithmic structures, and data requirements. We then systematically analyze their interplay, highlighting frameworks that integrate SFT and RL, hybrid training pipelines, and methods that leverage their complementary strengths. Drawing on a representative set of recent application studies from 2023 to 2025, we identify emerging trends, characterize the rapid shift toward hybrid post-training paradigms, and distill key takeaways that clarify when and why each method is most effective. By synthesizing theoretical insights, practical methodologies, and empirical evidence, this study establishes a coherent understanding of SFT and RL within a unified framework and outlines promising directions for future research in scalable, efficient, and generalizable LLM post-training.

cs.AI↗
arXiv2026

Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills

Large language model (LLM) agents are moving beyond prompting alone. ChatGPT marked the rise of general-purpose LLM assistants, DeepSeek showed that on-policy reinforcement learning with verifiable rewards can improve reasoning and tool use, and OpenClaw highlights a newer direction in which agents accumulate persistent memory and reusable skills. Yet the research landscape remains fragmented across post-training, retrieval, memory, and skill systems. This survey studies these developments under a single notion of \emph{adaptation}: improving an agent, its tools, or their interaction after pretraining. We organize the field with a four-paradigm framework spanning agent adaptation and tool adaptation. On the agent side, A1 (tool-execution-signaled) and A2 (agent-output-signaled) improve the agent itself through supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. On the tool side, T1 (agent-agnostic) provides reusable pre-trained modules any agent can call, while T2 (agent-supervised) uses the agent's outputs to train memory systems, skill libraries, or lightweight subagents. Using this framework, we review post-training methods, adaptive memory architectures, and agent skills; compare their trade-offs in cost, flexibility, and generalization; and summarize evaluation practices across deep research, software development, computer use, and drug discovery. We conclude by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.

cs.AI↗
arXiv2026

Frequency Security-Aware Production Scheduling of Utility-Scale Off-Grid Renewable P2H Systems Coordinating Heterogeneous Electrolyzers

Renewable power-to-hydrogen (ReP2H) enables large-scale renewable energy utilization and supports the decarbonization of hard-to-abate sectors, such as chemicals and maritime transport, via hydrogen-based renewable ammonia and methanol fuels. As a result, utility-scale ReP2H projects are expanding worldwide. However, off-grid ReP2H systems exhibit low inertia due to their converter-dominated nature, making frequency security a critical concern. Although recent studies show that electrolyzers can contribute to frequency regulation (FR), their support capability depends on operating states and loading levels, creating a trade-off between hydrogen output and frequency security. To address this challenge, this work develops a unified co-optimization framework for frequency security-aware production scheduling of utility-scale off-grid ReP2H systems coordinating heterogeneous electrolyzers. A system-level frequency response model is established to capture multi-stage FR from alkaline water electrolyzers (AWEs), proton exchange membrane electrolyzers (PEMELs), and other resources, including ammonia-fueled generators retrofitted in co-located chemical plants, battery energy storage, and wind turbines (WTs). Stage-wise transient frequency security constraints are derived, reformulated into tractable forms, and embedded into production scheduling, enabling coordinated on/off switching and load allocation across electrolyzers to maximize hydrogen output under uncertain renewable power input while enforcing frequency security constraints. Case studies based on real-world systems demonstrate that the proposed approach allows HPs to replace 55.52% and 96.85% of FR reserves from WTs and AFGs, respectively, while maintaining comparable hydrogen output. Year-long simulations show an average 28.96% increase in annual net profit resulting from reduced reliance on conventional reserves.

math.OC↗
arXiv2026

Effective running coupling constant and jet quenching parameter in the spinning background from holography

In this work, we study the effective running coupling constant of heavy quark pair and jet quenching parameter in the spinning background. Ultra-locally, the boosted fluid is described by the boosted parameter and dual to a globally rotating system. Our results show that the angular momentum suppresses the effective running coupling constant and reduces its maximum value. The results demonstrate that the angular momentum promotes the dissociation of quarkonium and has a stronger effect on the effective running coupling constant when the axis of $Q\overline{Q}$ is transverse to the direction of the angular momentum. We also find that the angular momentum enhances the jet quenching parameter and has a stronger effect when the jet moves transversely to the direction of the angular momentum, namely $\hat{q}_{\perp}> \hat{q}_{\parallel}$. We discuss the dependence of the jet quenching parameter on the $η/s$ at strong coupling in the presence of the angular momentum.

hep-ph↗
arXiv2026

TherapyGym: Evaluating and Aligning Clinical Fidelity and Safety in Therapy Chatbots

Large language models (LLMs) are increasingly used for mental-health support; yet prevailing evaluation methods--fluency metrics, preference tests, and generic dialogue benchmarks--fail to capture the clinically critical dimensions of psychotherapy. We introduce THERAPYGYM, a framework that evaluates and improves therapy chatbots along two clinical pillars: fidelity and safety. Fidelity is measured using the Cognitive Therapy Rating Scale (CTRS), implemented as an automated pipeline that scores adherence to CBT techniques over multi-turn sessions. Safety is assessed using a multi-label annotation scheme, covering therapy-specific risks (e.g., failing to address harm or abuse). To mitigate bias and unreliability in LLM-based judges, we further release THERAPYJUDGEBENCH, a validation set of 116 dialogues with 1,270 expert ratings for auditing and calibration against licensed clinicians. THERAPYGYM also serves as a training harness: CTRS and safety-based rewards drive RL with configurable patient simulations spanning diverse symptom profiles. Models trained in THERAPYGYM improve on expert ratings, with average CTRS rising from 0.10 to 0.60 (and 0.16 to 0.59 under LLM judges). Our work enables scalable development of therapy chatbots that are faithful to evidence-based practice and safer in high-stakes use.

cs.CL↗
arXiv2026

Correct Reasoning Paths Visit Shared Decision Pivots

Chain-of-thought (CoT) reasoning exposes the intermediate thinking process of large language models (LLMs), yet verifying those traces at scale remains unsolved. In response, we introduce the idea of decision pivots-minimal, verifiable checkpoints that any correct reasoning path must visit. We hypothesize that correct reasoning, though stylistically diverse, converge on the same pivot set, while incorrect ones violate at least one pivot. Leveraging this property, we propose a self-training pipeline that (i) samples diverse reasoning paths and mines shared decision pivots, (ii) compresses each trace into pivot-focused short-path reasoning using an auxiliary verifier, and (iii) post-trains the model using its self-generated outputs. The proposed method aligns reasoning without ground truth reasoning data or external metrics. Experiments on standard benchmarks such as LogiQA, MedQA, and MATH500 show the effectiveness of our method.

cs.AI↗
arXiv2026

Parabolic Anderson Model in the Hyperbolic Space. Part II: Quenched Asymptotics

We establish the exact quenched asymptotic growth of the solution to the parabolic Anderson model (PAM) in the hyperbolic space with a regular, stationary, time-independent Gaussian potential. More precisely, we show that with probability one, the solution $u$ to PAM with constant initial data has pointwise growth asymptotics \[ u(t,x)\sim e^{L^{*}t^{5/3}+o(t^{5/3})} \] as $t \rightarrow +\infty$. Both the power $t^{5/3}$ on the exponential and the exact value of $L^*$ are different from their counterparts in the Euclidean situation. They are determined through an explicit optimisation procedure. Our proof relies on certain fine localisation techniques, which also reveals a stronger non-Euclidean localisation mechanism.

math.PR↗
arXiv2026

Updatable Balanced Index for Stable Streaming Similarity Search over Large-Scale Fresh Vectors

As artificial intelligence gains more and more popularity, vectors are one of the most widely used data structures for services such as information retrieval and recommendation. Approximate Nearest Neighbor Search (ANNS), which generally relies on indices optimized for fast search to organize large datasets, has played a core role in these popular services. As the frequency of data shift grows, it is crucial for indices to accommodate new data and support real-time updates. Existing researches adopting two different approaches hold the following drawbacks: 1) approaches using additional buffers to temporarily store new data are resource-intensive and inefficient due to the global rebuilding processes; 2) approaches upgrading the internal index structure suffer from performance degradation because of update congestion and imbalanced distribution in streaming workloads. In this paper, we propose UBIS, an Updatable Balanced Index for stable streaming similarity Search, to resolve conflicts by scheduling concurrent updates and maintain good index quality by reducing imbalanced update cases, when the update frequency grows. Experimental results in the real-world datasets demonstrate that UBIS achieves up to 77% higher search accuracy and 45% higher update throughput on average compared to the state-of-the-art indices in streaming workloads.

cs.DB↗
arXiv2026

Equilibrium Investment with Random Risk Aversion: (Non-)uniqueness, Optimality, and Comparative Statics

This paper studies a continuous-time portfolio selection problem under a general distribution of random risk aversion (RRA). We provide a complete characterization of all deterministic equilibrium strategies in closed form. Our results show that the structure of the solution depends crucially on the distribution of RRA: the equilibrium is unique (if exits) when the expectation of RRA is finite, whereas an infinite expectation leads either to infinitely many equilibria or to a unique trivial one (i.e. risk-free investment). To resolve this multiplicity of equilibria, we select, among all deterministic equilibria, the one that maximizes the objective functional at the initial time. We establish a necessary and sufficient condition for the existence of such an optimal equilibrium, which is then shown to be unique and uniformly optimal. Finally, we conduct a comparative statics. Using counterexamples based on two-point distributed RRA, we demonstrate that a larger risk aversion in the sense of first-order stochastic dominance does not necessarily lead to less risky investment. Within the two-point distribution framework, we further examine the single-crossing property of equilibrium strategies and the monotonicity of the crossing time. We show that a larger risk aversion under a stronger stochastic order -- the reverse hazard rate order -- always leads to less risky investment. In addition, we analyze how the convex combination of independent and identically distributed RRAs influences investment.

q-fin.MF↗