arXiv subjects

Dawei Xiang

Publications and source records attributed to Dawei Xiang.

At least 19 recordsLinked to original sources

arXiv2026

FactorSpec: Serving Exact Relational Constraints as Speculative Drafts

Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities' stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.

cs.LG↗

arXiv2026

Repair Changes Which Layers to Remove: Exact Optima for Depth Pruning of Language Models

Depth pruning removes entire transformer layers to make large language models faster, and recent methods recover quality by replacing the removed layers with a training-free linear map. Which layers to remove, however, is chosen with heuristic scores whose distance from the best choice is unknown. We evaluate every subset of layer groups exactly (about 53,000 pruned models across four families, with and without repair) and find that the best layers to remove depend on the repair: linear repair absorbs the interaction among nearby layers and turns the optimum from a split layout into a contiguous block, so methods that choose layers without the repair miss the repaired optimum by up to 189%. We therefore select layers under the repair that will be applied. A greedy search at the cost of SLEB lowers perplexity by 20% to 66% relative to ReplaceMe on five models up to 27B parameters, and with a task objective it is up to 25 points more accurate given the same data. A second-order expansion of the loss yields a score that selects the repaired block from one forward and one backward pass of the intact model.

cs.DC↗

arXiv2026

Stronger Judges Do Not See More: Deployment-Integrity Blindness in LLM Evaluation

LLM-agent evaluation reports behavioral scores but leaves the identity of the deployed runtime stack unreported. A tampered deployment that preserves evaluated outputs passes every output-only evaluator. We introduce DIB-Bench, a benchmark that measures this deployment-integrity blindness across 7,560 scored examples, six tamper types, and an evaluator ladder from regex scorers through LLM judges to GPT-5-nano. Stronger evaluators accept more tampered outputs: mean acceptance rises from 88.6% (rule-based) to 94.1% (GPT-5-nano), and a trigger-conditional tamper invisible on the evaluation suite produces 48-100% harmful outputs, rising with model capability. A PAC-style bound shows that tampers with zero output shift are undetectable from outputs at any sample size; a manifest-hash identity channel detects them with zero samples. Behavioral fingerprinting identifies the model component above chance (76.7%) but covers one of six stack components. We also propose a graded deployment-identity descriptor for evaluation reports.

cs.CR↗

arXiv2026

Dynamic Expert Quantization for Scalable Mixture-of-Experts Inference

Mixture-of-Experts (MoE) has become a practical architecture for scaling LLM capacity while keeping per-token compute modest, but deploying MoE models on a single, memory-limited GPU remains difficult because expert weights dominate the HBM footprint. Existing expert offloading and prefetching systems reduce the resident set, yet they often pay expert-loading costs on the critical path when activation becomes dense. Post-training quantization (PTQ) lowers the footprint without transfers, but prevailing pipelines fix expert bit-widths offline and assume routing remains stable, even though MoE expert utilization is heavy-tailed and the hot set can shift across workloads. We present DynaExq, a runtime-aware mixed-precision serving system that treats single-GPU MoE inference under a hard HBM envelope as an online, budget-constrained precision allocation problem. The key insight is to keep the experts that dominate runtime traffic resident at higher precision, while maintaining a low-precision fallback for the remaining experts, so the system can reduce transfer volume and avoid the waiting latency that limits offloading and prefetching under dense activation. DynaExq estimates long-horizon expert hotness from router traces, selects a per-layer high-precision resident set via a budget-feasible top-$n$ rule, and applies promotions and demotions asynchronously through stable expert handles so the forward pass always executes on a fully materialized expert version. Across Qwen3-MoE-30B/80B and six benchmarks, DynaExq improves accuracy over static PTQ on Qwen3-80B (73.09% to 77.57%) under comparable device-memory budgets and achieves up to 2.73x higher throughput than offloading/prefetch baselines at batch size 32.

cs.PF↗

arXiv2026

Latency-Quality Routing for Functionally Equivalent Tools in LLM Agents

Tool-augmented LLM agents increasingly access the same tool type through multiple functionally equivalent providers, such as web-search APIs, retrievers, or LLM backends exposed behind a shared interface. This creates a provider-routing problem under runtime load: the router must choose among providers that differ in latency, reliability, and answer quality, often without gold labels at deployment time. We introduce LQM-ContextRoute, a contextual bandit router for same-function tool providers. Its key design is latency-quality matching: instead of letting low latency offset poor answers in an additive reward, the router ranks providers by expected answer quality per service cycle. It combines this capacity-aware score with query-specific quality estimation and LLM-as-judge feedback, allowing it to adapt online to both load changes and provider-quality differences. On the main web-search load benchmark, LQM-ContextRoute improves F1 by +2.18 pp over SW-UCB while staying on the latency-quality frontier. In a high-heterogeneity StrategyQA setting, LQM-ContextRoute avoids additive-reward collapse and improves accuracy by up to +18 pp over SW-UCB; on heterogeneous retriever pools, it improves NDCG by +2.91--+3.22 pp over SW-UCB. These results show that same-function tool routing benefits from treating latency as service capacity, especially when runtime pressure and provider-quality heterogeneity coexist.

cs.LG↗

arXiv2026

Selective KV-Cache Sharing to Mitigate Timing Side-Channels in LLM Inference

Global KV-cache sharing is an effective optimization for accelerating large language model (LLM) inference, yet it introduces an API-visible timing side channel that lets adversaries infer sensitive user inputs from shared entries, leading to cross-tenant privacy risks. To address this problem, we introduce SafeKV (Secure and Flexible KV-cache Sharing), a system-level co-design of privacy enforcement and KV-cache management. SafeKV integrates lightweight detection and isolation directly into the serving runtime to eliminate cross-tenant reuse of sensitive KV-cache blocks under our threat model, while recovering most of the performance benefits of global sharing. Our key contributions are: (1) a three-tier asynchronous detection pipeline that decouples privacy classification from inference and supports streaming workloads, (2) a unified radix-tree-based memory manager with path compression and sensitivity-aware eviction for scalable selective isolation, and (3) an RDR-guided (Reuse Diversity Ratio) runtime safeguard that detects and bounds residual leakage. On large LLM backends, SafeKV reduces the time-to-first-token (TTFT) overhead compared to full isolation by up to 40.58% and raises throughput by up to 2.66x. Overall, SafeKV restores the efficiency of KV reuse while enforcing strong, practical privacy for multi-tenant LLM inference.

cs.CR↗

arXiv2026

TreeDiff: AST-Guided Code Generation with Diffusion LLMs

Code generation is increasingly critical for real-world applications. Still, diffusion-based large language models continue to struggle with this demand. Unlike free-form text, code requires syntactic precision; even minor structural inconsistencies can render a program non-executable. Existing diffusion-based large language models rely on random token masking for corruption, leading to two key failures: they lack awareness of syntactic boundaries during the iterative denoising process, and they fail to capture the long-range hierarchical dependencies essential for program correctness. We propose TreeDiff to address both issues. Specifically, we propose a syntax-aware diffusion framework that incorporates structural priors from Abstract Syntax Tree (AST) into the corruption process. Instead of masking individual tokens at random, we selectively mask tokens belonging to key AST nodes. By aligning the corruption process with the underlying structure of code, our method encourages the model to internalize the compositional nature of programming languages, enabling it to reconstruct programs that respect grammatical boundaries and capture long-range dependencies. Our method achieves a 13.3% relative improvement over the random masking training method, demonstrating its effectiveness in code generation task by leveraging underlying structures.

cs.CL↗

arXiv2025

HyperEdit: Unlocking Instruction-based Text Editing in LLMs via Hypernetworks

Instruction-based text editing is increasingly critical for real-world applications such as code editors (e.g., Cursor), but Large Language Models (LLMs) continue to struggle with this task. Unlike free-form generation, editing requires faithfully implementing user instructions while preserving unchanged content, as even minor unintended modifications can break functionality. Existing approaches treat editing as generic text generation, leading to two key failures: they struggle to faithfully align edits with diverse user intents, and they often over-edit unchanged regions. We propose HyperEdit to address both issues. First, we introduce hypernetwork-based dynamic adaptation that generates request-specific parameters, enabling the model to tailor its editing strategy to each instruction. Second, we develop difference-aware regularization that focuses supervision on modified spans, preventing over-editing while ensuring precise, minimal changes. HyperEdit achieves a 9%--30% relative improvement in BLEU on modified regions over state-of-the-art baselines, despite utilizing only 3B parameters.

cs.CL↗

arXiv2025

ExpertFlow: Adaptive Expert Scheduling and Memory Coordination for Efficient MoE Inference

The expansion of large language models is increasingly limited by the constrained memory capacity of modern GPUs. To mitigate this, Mixture-of-Experts (MoE) architectures activate only a small portion of parameters during inference, significantly lowering both memory demand and computational overhead. However, conventional MoE inference approaches, which select active experts independently at each layer, often introduce considerable latency because of frequent parameter transfers between host and GPU memory. In addition, current cross-layer prediction strategies, which are typically based on fixed steps, lack adaptability across different hardware platforms and workloads, thereby reducing their robustness and effectiveness. To address these challenges, we present ExpertFlow, a runtime system for MoE inference that combines adaptive expert prefetching and cache-aware routing. ExpertFlow continuously adjusts its prediction horizon for expert activation by leveraging runtime statistics such as transfer bandwidth, parameter dimensionality, and model feedback signals. Furthermore, it incorporates a hybrid cross-layer prediction scheme that fuses pregating information with intermediate computational states to anticipate future expert needs. By adaptively refining prefetching decisions and aligning them with actual usage behavior, ExpertFlow effectively decreases cache misses and removes latency caused by expert swap-ins. Our evaluation demonstrates that ExpertFlow reduces model stall time to less than 0.1% of the baseline, highlighting its capability to optimize MoE inference under stringent memory constraints.

cs.DC↗

arXiv2025

MMM-Fact: A Multimodal, Multi-Domain Fact-Checking Dataset with Multi-Level Retrieval Difficulty

Misinformation and disinformation demand fact checking that goes beyond simple evidence-based reasoning. Existing benchmarks fall short: they are largely single modality (text-only), span short time horizons, use shallow evidence, cover domains unevenly, and often omit full articles -- obscuring models' real-world capability. We present MMM-Fact, a large-scale benchmark of 125,449 fact-checked statements (1995--2025) across multiple domains, each paired with the full fact-check article and multimodal evidence (text, images, videos, tables) from four fact-checking sites and one news outlet. To reflect verification effort, each statement is tagged with a retrieval-difficulty tier -- Basic (1--5 sources), Intermediate (6--10), and Advanced (>10) -- supporting fairness-aware evaluation for multi-step, cross-modal reasoning. The dataset adopts a three-class veracity scheme (true/false/not enough information) and enables tasks in veracity prediction, explainable fact-checking, complex evidence aggregation, and longitudinal analysis. Baselines with mainstream LLMs show MMM-Fact is markedly harder than prior resources, with performance degrading as evidence complexity rises. MMM-Fact offers a realistic, scalable benchmark for transparent, reliable, multimodal fact-checking.

cs.SI↗

arXiv2025

PromptSculptor: Multi-Agent Based Text-to-Image Prompt Optimization

The rapid advancement of generative AI has democratized access to powerful tools such as Text-to-Image models. However, to generate high-quality images, users must still craft detailed prompts specifying scene, style, and context-often through multiple rounds of refinement. We propose PromptSculptor, a novel multi-agent framework that automates this iterative prompt optimization process. Our system decomposes the task into four specialized agents that work collaboratively to transform a short, vague user prompt into a comprehensive, refined prompt. By leveraging Chain-of-Thought reasoning, our framework effectively infers hidden context and enriches scene and background details. To iteratively refine the prompt, a self-evaluation agent aligns the modified prompt with the original input, while a feedback-tuning agent incorporates user feedback for further refinement. Experimental results demonstrate that PromptSculptor significantly enhances output quality and reduces the number of iterations needed for user satisfaction. Moreover, its model-agnostic design allows seamless integration with various T2I models, paving the way for industrial applications.

cs.MA↗

arXiv2025

MVVM: Deploy Your AI Agents-Securely, Efficiently, Everywhere

The rise of AI agents powered by Large Language Models (LLMs) presents critical challenges: how to securely execute and migrate these agents across heterogeneous environments while protecting sensitive user data, maintaining availability during network failures, minimizing response latency for time-critical decisions, and ensuring output safety in mission-critical applications. We present MVVM, a WebAssembly-based secure container framework that enables transparent live migration of LLM agent workspaces between edge devices and cloud servers with end-to-end privacy guarantees, resilient multi-tier replication, speculative execution for latency optimization, and integrated validation for safety assurance. MVVM introduces two key innovations: (1) a two-way sandboxing framework leveraging hardware enclaves and accelerator extensions that protects both the agent from malicious hosts and the host from compromised agents; (2) an efficient cross platform migration mechanism using WebAssembly and WASI's platform-agnostic design, enabling seamless movement across ARM phones, RISC-V MCUs, x86 servers, and heterogeneous accelerators; and three astonishing use cases: (1) privacy-aware daemon that automatically determines whether to execute locally or remotely based on data sensitivity and resource availability; (2) multi-tier replication with intelligent quality degradation that maintains service availability despite network failures or resource constraints; (3) a comprehensive execution framework combining speculative execution for 10x latency reduction with parallel validation that ensures output safety without compromising responsiveness. Our evaluation demonstrates that MVVM is validated on three separate devices across 18 workloads.

cs.OS↗

arXiv2025

Mining Intraday Risk Factor Collections via Hierarchical Reinforcement Learning based on Transferred Options

Traditional risk factors like beta, size/value, and momentum often lag behind market dynamics in measuring and predicting stock return volatility. Statistical models like PCA and factor analysis fail to capture hidden nonlinear relationships. Genetic programming (GP) can identify nonlinear factors but often lacks mechanisms for evaluating factor quality, and the resulting formulas are complex. To address these challenges, we propose a Hierarchical Proximal Policy Optimization (HPPO) framework for automated factor generation and evaluation. HPPO uses two PPO models: a high-level policy assigns weights to stock features, and a low-level policy identifies latent nonlinear relationships. The Pearson correlation between generated factors and return volatility serves as the reward signal. Transfer learning pre-trains the high-level policy on large-scale historical data, fine-tuning it with the latest data to adapt to new features and shifts. Experiments show the HPPO-TO algorithm achieves a 25\% excess return in HFT markets across China (CSI 300/800), India (Nifty 100), and the US (S\&P 500). Code and data are available at https://github.com/wencyxu/HRL-HF_risk_factor_set.

cs.CE↗

arXiv2025

Modality-Aware Infrared and Visible Image Fusion with Target-Aware Supervision

Infrared and visible image fusion (IVIF) is a fundamental task in multi-modal perception that aims to integrate complementary structural and textural cues from different spectral domains. In this paper, we propose FusionNet, a novel end-to-end fusion framework that explicitly models inter-modality interaction and enhances task-critical regions. FusionNet introduces a modality-aware attention mechanism that dynamically adjusts the contribution of infrared and visible features based on their discriminative capacity. To achieve fine-grained, interpretable fusion, we further incorporate a pixel-wise alpha blending module, which learns spatially-varying fusion weights in an adaptive and content-aware manner. Moreover, we formulate a target-aware loss that leverages weak ROI supervision to preserve semantic consistency in regions containing important objects (e.g., pedestrians, vehicles). Experiments on the public M3FD dataset demonstrate that FusionNet generates fused images with enhanced semantic preservation, high perceptual quality, and clear interpretability. Our framework provides a general and extensible solution for semantic-aware multi-modal image fusion, with benefits for downstream tasks such as object detection and scene understanding.

cs.CV↗

arXiv2025

FinMultiTime: A Four-Modal Bilingual Dataset for Financial Time-Series Analysis

Pure time series forecasting tasks typically focus exclusively on numerical features; however, real-world financial decision-making demands the comparison and analysis of heterogeneous sources of information. Recent advances in deep learning and large scale language models (LLMs) have made significant strides in capturing sentiment and other qualitative signals, thereby enhancing the accuracy of financial time series predictions. Despite these advances, most existing datasets consist solely of price series and news text, are confined to a single market, and remain limited in scale. In this paper, we introduce FinMultiTime, the first large scale, multimodal financial time series dataset. FinMultiTime temporally aligns four distinct modalities financial news, structured financial tables, K-line technical charts, and stock price time series across both the S&P 500 and HS 300 universes. Covering 5,105 stocks from 2009 to 2025 in the United States and China, the dataset totals 112.6 GB and provides minute-level, daily, and quarterly resolutions, thus capturing short, medium, and long term market signals with high fidelity. Our experiments demonstrate that (1) scale and data quality markedly boost prediction accuracy; (2) multimodal fusion yields moderate gains in Transformer models; and (3) a fully reproducible pipeline enables seamless dataset updates.

cs.CE↗

arXiv2025

Learning Explainable Stock Predictions with Tweets Using Mixture of Experts

Stock price movements are influenced by many factors, and alongside historical price data, tex-tual information is a key source. Public news and social media offer valuable insights into market sentiment and emerging events. These sources are fast-paced, diverse, and significantly impact future stock trends. Recently, LLMs have enhanced financial analysis, but prompt-based methods still have limitations, such as input length restrictions and difficulties in predicting sequences of varying lengths. Additionally, most models rely on dense computational layers, which are resource-intensive. To address these challenges, we propose the FTS- Text-MoE model, which combines numerical data with key summaries from news and tweets using point embeddings, boosting prediction accuracy through the integration of factual textual data. The model uses a Mixture of Experts (MoE) Transformer decoder to process both data types. By activating only a subset of model parameters, it reduces computational costs. Furthermore, the model features multi-resolution prediction heads, enabling flexible forecasting of financial time series at different scales. Experimental results show that FTS-Text-MoE outperforms baseline methods in terms of investment returns and Sharpe ratio, demonstrating its superior accuracy and ability to predict future market trends.

cs.CE↗

arXiv2025

AI-Driven Prognostics for State of Health Prediction in Li-ion Batteries: A Comprehensive Analysis with Validation

This paper presents a comprehensive review of AI-driven prognostics for State of Health (SoH) prediction in lithium-ion batteries. We compare the effectiveness of various AI algorithms, including FFNN, LSTM, and BiLSTM, across multiple datasets (CALCE, NASA, UDDS) and scenarios (e.g., varying temperatures and driving conditions). Additionally, we analyze the factors influencing SoH fluctuations, such as temperature and charge-discharge rates, and validate our findings through simulations. The results demonstrate that BiLSTM achieves the highest accuracy, with an average RMSE reduction of 15% compared to LSTM, highlighting its robustness in real-world applications.

cs.AI↗

arXiv2025

Neural Pruning for 3D Scene Reconstruction: Efficient NeRF Acceleration

Neural Radiance Fields (NeRF) have become a popular 3D reconstruction approach in recent years. While they produce high-quality results, they also demand lengthy training times, often spanning days. This paper studies neural pruning as a strategy to address these concerns. We compare pruning approaches, including uniform sampling, importance-based methods, and coreset-based techniques, to reduce the model size and speed up training. Our findings show that coreset-driven pruning can achieve a 50% reduction in model size and a 35% speedup in training, with only a slight decrease in accuracy. These results suggest that pruning can be an effective method for improving the efficiency of NeRF models in resource-limited settings.

cs.CV↗