arXiv subjects

Wei Shao

Publications and source records attributed to Wei Shao.

At least 19 recordsLinked to original sources

arXiv2026

AMEND: Audited Margins Enable Nonblocking Drops in GPU-PIM LLM Decoding

Autoregressive large language model (LLM) decoding re-reads a growing key-value (KV) cache at every step, so long-context attention is bound by graphics processing unit (GPU) memory bandwidth. Block-sparse attention skips low-contribution KV blocks, but a selector that decides after the current query-key (QK) product, such as max-relative block thresholding (BLASST), still reads every K block, and a processing-in-memory (PIM) filter that decides from the current query places a serial PIM stage on the critical path. We present AMEND, a GPU-PIM attention design that removes both dependencies. AMEND predicts each block's BLASST verdict from margins audited at earlier steps, so the GPU fetches only predicted survivors while near-bank PIM units in high-bandwidth memory (HBM) concurrently score the omitted complement. A stack-level controller merges both observations, updates the predictor, and eagerly generates the next step's mask, so every predicted drop is re-observed without blocking the current token. Operating points are selected offline by constrained Bayesian optimization under a false-drop budget. Across LongBench and RULER runs, AMEND preserves near-baseline task quality; in simulation at batch size 8, it achieves $1.40$-$3.63\times$ end-to-end decode speedup and 28-66% lower dynamic decode energy than dense attention across 8K-64K contexts.

cs.AR↗

arXiv2026

Multi-Grid Post-Training for Long-Form Multi-Shot Video Generation

Generating long-form multi-shot videos requires coherent within-shot motion and visually consistent narratives across shots. Existing video generators favor continuous motion and struggle to present complete shot sets when an entire narrative is packed along one temporal axis. We propose MovieGrid, a Multi-Grid Post-Training paradigm that decomposes a long video into shorter, temporally ordered chunks and arranges them on a spatial grid for joint modeling. This design reduces the number of shots handled by each temporal axis while enabling global information exchange across chunks. We construct the Multi-Grid Long Video (MGLV) dataset from 1,000 long-form videos using source video collection, hierarchical segmentation, grid video construction, and character-aware story annotation, producing 54K grid videos paired with story prompts. Our Noise-Free Random-Grid Training retains a random subset of chunks as clean visual context for denoising the remaining chunks. Grid Embedding encodes grid structure, character-aware Story Prompts link recurring entities, and Grid Boundary Loss stabilizes layouts. Under the same token budget, MovieGrid generates 6.05 times more shots than Temporal Packing in a 1,616-frame video. On a benchmark spanning five real-world categories, it achieves state-of-the-art intra-shot consistency (0.9131 versus 0.8086 for HoloCine) and inter-shot consistency (0.5914 versus 0.5384 for StoryMem). MovieGrid can further scale video length with minimal compromise through single or multiple generations.

cs.CV↗

arXiv2026

Beyond End-to-End Success: Diagnosing Failures in Long-Horizon Security LLM Agents

Long-horizon security LLM agents must carry information and decisions across many dependent interactions, where later actions often depend on services, state, or access discovered much earlier. This makes final task success difficult to interpret: an agent may fail before it ever reaches the point where the capability of interest can be exercised. We present a diagnostic methodology that instruments security tasks with checkpoints, separates failures before and after capability exposure, and uses controlled interventions to test suspected upstream bottlenecks. We evaluate the methodology across four task families involving delayed reuse of discovered information, reuse of observed state, recovery from failed strategies, and decision making after uncertain outcomes. On observed state reuse, checkpoint analysis shows that many Gemini 2.5 Flash failures occur before the model observes the state it is later expected to reuse. In a pre-specified 92-seed study, targeted protocol-disambiguation guidance increases state observation from 65.5\% under a matched non-guidance control message to 95.4\%. Repeating the same design with Gemini 3.7 Flash produces the opposite effect, while state observation no longer reliably predicts task completion. These results show that the dominant source of failure can shift across model generations, motivating evaluation that diagnoses where and why long-horizon security agents fail rather than relying only on aggregate task success.

cs.CR↗

arXiv2026

SpO$_2$ Predictor-Guided Stage-Wise Time-Frequency Reconstruction of Low-Quality Dual-Wavelength PPG for Oxygen Saturation Estimation

Continuous oxygen saturation (SpO$_2$) estimation from wearable photoplethysmography (PPG) is important for long-term health monitoring, but low-quality red and infrared PPG segments can distort waveform morphology and degrade SpO$_2$ prediction accuracy. Existing PPG denoising and reconstruction methods usually optimize waveform fidelity or heart rate characteristics, while time-domain waveform loss on PPG signals alone insufficiently preserves frequency structure and SpO$_2$-relevant information. This paper proposes a SpO$_2$ predictor-guided stage-wise time-frequency reconstruction framework for low-quality dual-wavelength PPG signals. The proposed method first selects high-quality PPG segments to pretrain a SpO$_2$ predictor. A masked reconstruction model is then trained to recover randomly masked PPG regions using a joint reconstruction objective that combines time-domain waveform loss with frequency-domain loss computed from the short-time Fourier transform (STFT). To make the reconstruction task physiologically relevant, the pretrained SpO$_2$ predictor is incorporated as an additional constraint, encouraging the reconstructed PPG to preserve SpO$_2$ information rather than only minimizing waveform reconstruction error. The SpO$_2$ predictor and PPG reconstructor model are optimized through four training stages. Experiments on the public OpenOximetry Repository and a private wearable PPG dataset show that the proposed approach achieves the lowest subject-level MAE, with 2.882\% on the public dataset and 2.359\% on the private dataset.

eess.SP↗

arXiv2026

Low-Rate Wrist SpO2 Estimation under Micro-Perturbations Using Motion-Aware Beat Selection and Perfusion-Guided Calibration

Continuous oxygen saturation (SPO2) monitoring from photoplethysmography (PPG) is important for wearable health sensing, but wrist-based SPO2 estimation remains challenging due to subtle wrist micro-perturbations and inter-subject differences in local perfusion status. These factors can destabilize the red-to-infrared ratio-of-ratios (R) and reduce the reliability of conventional fixed R-SPO2 mapping. This paper proposes a lightweight low-rate wrist SPO2 estimation framework that integrates motion-aware beat selection and perfusion-guided calibration. The proposed method extracts beat-level alternating-current/direct-current (AC/DC) components from dual-wavelength PPG signals, computes beat-level R values, and uses accelerometer-derived motion scores to weight beats within each sliding window. A subject-specific perfusion reference is further used to guide calibration across different perfusion conditions. Experiments on a private wearable dataset show that the proposed method achieves the best 25 Hz performance, with an MAE of 2.305$\pm$1.113% and an RMSE of 3.117$\pm$1.743%, while maintaining performance comparable to the 100 Hz sampling rate and reducing PPG sensor power consumption for energy-efficient wearable implementation. These results demonstrate the effectiveness of the proposed framework for low-rate wrist SPO2 estimation under micro-perturbations.

eess.SP↗

arXiv2026

Unit-Independent Low-Rate Wrist GSR Processing for Stress Detection Using Phasic nSCR Features

Galvanic skin response (GSR) is widely used for stress detection, but wrist-based GSR remains challenging because its absolute amplitude can differ substantially from laboratory-grade palmar measurements. In this paper, we propose a unit-independent low-rate wrist GSR processing pipeline to extract the number of skin conductance responses per minute (nSCR/min) as a stress-related feature. We collect paired wrist and palmar GSR recordings from 31 participants during sitting baseline, standing baseline, neutral speaking, and the Trier Social Stress Test (TSST), a laboratory social stressor task. The proposed pipeline cleans the raw GSR signal, decomposes it into tonic skin conductance level (SCL) and phasic skin conductance response (SCR), applies robust z-score normalization, and detects phasic SCR peaks to compute nSCR/min. Using random forest on 25Hz We-Be GSR, nSCR/min achieved balanced accuracies of 0.823 and 0.871 for binary classification between TSST and the sitting and standing baselines, respectively. Moreover, the 25Hz We-Be GSR features achieved comparable balanced accuracy to the original 100Hz features across the evaluated tasks. These results suggest the feasibility of low-rate, unit-independent wrist GSR processing for wearable stress detection.

eess.SP↗

arXiv2026

Generation Quality-Latency Tradeoff-Aware Inference Offloading for Multimodal LLMs in Cloud-Edge Continuum

Beyond pure cloud, some efforts are being made to deploy Large Language Models (LLMs) in edge to accelerate inference response. So the deployment of LLMs in cloud-edge continuum becomes a promising paradigm, where the tasks involving multimodal data occupy a large part of requests. Under this continuum, users usually concern about multiple Quality-of-Service (QoS) attributes, but it is always intractable to jointly optimize them. In this paper, we propose to study the joint optimization of those attributes and focus on two key representatives, i.e., content generation quality and response latency. We propose to study the offloading technology to achieve a tradeoff between the two objectives in the cloud-edge collaborative Multimodal LLM (MLLM) system. However, it is highly difficult to predict generation quality and inference latency for MLLM inference tasks while optimizing this offloading process. To address these unprecedented difficulties, we propose a Quality-Latency Tradeoff-Aware MLLM Inference Offloading (QLMIO) framework to make decisions that optimally balance generation quality and response latency. Meanwhile, recognizing the absence of publicly available datasets tailored to the MLLM inference offloading problem, we constructed a real-world cloud-edge collaborative MLLM system and subsequently collected an MLLM Inference Offloading Benchmark (MIOBench) to comprehensively evaluate our framework and facilitate the study of this problem. Extensive experimental results demonstrate that the QLMIO framework reduces latency by up to 58.14\% compared to baselines, while simultaneously matching the task completion rate achieved under the case that executes all requests exclusively on a cloud server. The dataset and codes are available at Github.

cs.DC↗

arXiv2026

Lightweight Cross-Device Sleep Tracking on the WeBe Wearable Platform

Wearable devices are widely used for continuous health monitoring, yet reliable sleep tracking on emerging platforms remains underexplored due to reliance on proprietary algorithms and device-specific activity representations. We present a lightweight and reproducible sleep tracking pipeline that operates directly on raw accelerometer signals. The method converts data into epoch-level activity features, applies temporal smoothing and normalized scoring, and performs sleep/wake classification using a globally calibrated threshold. We calibrate the model on the Multilevel Monitoring of Activity and Sleep in Healthy People (MMASH) dataset and evaluate it in a cross-device study using the WeBe wearable platform and a commercial ActiGraph device. On MMASH, the method achieves a mean absolute error of 41.6 minutes in Total Sleep Time (TST), with onset and offset errors of 6.3 and 7.4 minutes. On real-world WeBe data from three participants across five sessions, it achieves a mean TST error of 27.4 minutes and onset and offset errors of 13.9 and 8.0 minutes. In contrast, a commercial ActiGraph pipeline shows larger discrepancies relative to ground truth. These results demonstrate accurate and generalizable sleep tracking using a simple and reproducible pipeline.

cs.ET↗

arXiv2026

AIS: Adaptive Importance Sampling for Quantized RL

Reinforcement learning (RL) for large language models (LLMs) is dominated by the cost of rollout generation, which has motivated the use of low-precision rollouts (e.g., FP8) paired with a BF16 trainer to improve throughput and reduce memory pressure. This introduces a rollout-training mismatch that biases the policy gradient and can cause training to collapse outright on reasoning benchmarks. We show that the mismatch is non-stationary and acts as a double-edged sword: early in training it provides a stochastic exploration bonus, exposing the gradient to trajectories the trainer would otherwise under-sample, but the same perturbation transitions into a destabilizing source of bias as the policy concentrates. To solve this, we propose Adaptive Importance Sampling (AIS), a correction framework that adjusts the strength of its intervention on a per-batch basis. AIS combines three real-time diagnostics, namely weight reliability, divergence severity, and variance amplification, into a single mixing coefficient that interpolates between the uncorrected and fully importance-weighted gradients, suppressing the destabilizing component of the mismatch while preserving its exploratory benefit. We integrate AIS into GRPO and evaluate it on the diffusion-based LLaDA-8B-Instruct and the autoregressive Qwen3-8B and Qwen3.5-9B across mathematical reasoning and planning benchmarks. AIS matches the BF16 baseline on most tasks while retaining the 1.5 to 2.76x rollout speedup of FP8.

stat.ML↗

arXiv2026

Multi-domain Multi-modal Document Classification Benchmark with a Multi-level Taxonomy

Document classification forms the backbone of modern enterprise content management, yet existing benchmarks remain trapped in oversimplified paradigms -- single domain settings with flat label structures -- that bear little resemblance to the hierarchical, multi-modal, and cross-domain nature of real-world business documents. This gap not only misrepresents practical complexity but also stifles progress toward industrially viable document intelligence. To bridge this gap, we construct the first Multi-level, Multi-domain, Multi-modal document classification Benchmark (MMM-Bench). MMM-Bench includes (1) a deeply hierarchical taxonomy spanning five levels that capture the authentic organizational logic of business documentation; and (2) 5,990 real-world multi-modal documents meticulously curated from 12 commercial domains in Alibaba. Each document is manually annotated with a complete hierarchical path by domain experts. We establish comprehensive baselines on MMM-Bench, which consists of open-weight models and API-based models. Through systematic experiments, we identify four fundamental challenges within MMM-Bench and propose corresponding insights. To provide a solid foundation for advancing research in multi-level, multi-domain document classification, we release all of the data and the evaluation toolkit of MMM-Bench at https://github.com/MMMDC-Bench/MMMDC-Bench.

cs.CL↗

arXiv2026

ESICA: A Scalable Framework for Text-Guided 3D Medical Image Segmentation

Text guided 3D medical image segmentation offers a flexible alternative to class based and spatial prompt based models by allowing users to specify regions of interest directly in natural language. This paradigm avoids reliance on predefined label sets, reduces ambiguous outputs, and aligns more naturally with clinical workflows. However, existing text guided frameworks are often computationally expensive, exhibit weak text volume feature alignment, and fail to capture fine anatomical details. We propose ESICA, a lightweight and scalable framework that addresses these challenges through three innovations: (1) a similarity matrix based mask prediction formulation that enhances semantic alignment, (2) an efficient decomposed decoder with adapter modules for accurate volumetric decoding, and (3) a two pass refinement strategy that sharpens boundaries and resolves uncertain regions. To improve training stability and generalization, ESICA adopts a two stage scheme consisting of positive only pretraining followed by balanced fine tuning. On the CVPR BiomedSegFM benchmark spanning five imaging modalities (CT, MRI, PET, ultrasound, and microscopy), ESICA achieves state of the art segmentation accuracy, while the compact ESICA4 Lite variant attains similar segmentation performance with substantially fewer parameters, yielding a superior efficiency accuracy trade off. Our framework advances text guided segmentation toward efficient, scalable, and clinically deployable systems. Code will be made publicly available at https://github.com/mirthAI/ESICA.

cs.CV↗

arXiv2026

Detoxification for LLM: From Dataset Itself

Existing detoxification methods for large language models mainly focus on post-training stage or inference time, while few tackle the source of toxicity, namely, the dataset itself. Such training-based or controllable decoding approaches cannot completely suppress the model's inherent toxicity, whereas detoxifying the pretraining dataset can fundamentally reduce the toxicity that the model learns during training. Hence, we attempt to detoxify directly on raw corpora with SoCD (Soft Contrastive Decoding), which guides an LLM to localize and rewrite toxic spans in raw data while preserving semantics, in our proposed HSPD (Hierarchical Semantic-Preserving Detoxification) pipeline, yielding a detoxified corpus that can drop-in replace the original for fine-tuning or other training. On GPT2-XL, HSPD attains state-of-the-art detoxification, reducing Toxicity Probability (TP) from 0.42 to 0.18 and Expected Maximum Toxicity (EMT) from 0.43 to 0.20. We further validate consistent best-in-class results on LLaMA2-7B, OPT-6.7B, and Falcon-7B. These findings show that semantics-preserving, corpus-level rewriting with HSPD effectively suppresses downstream toxicity while retaining data utility and allowing seamless source-level mitigation, thereby reducing the cost of later model behavior adjustment. (Code is available at: https://github.com/ntsw2001/data_detox_for_llm)

cs.CL↗

arXiv2026

Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement

Three-dimensional (3D) medical image enhancement, including denoising and super-resolution, is critical for clinical diagnosis in CT, PET, and MRI. Although diffusion models have shown remarkable success in 2D medical imaging, scaling them to high-resolution 3D volumes remains computationally prohibitive due to lengthy diffusion trajectories over high-dimensional volumetric data. We observe that in conditional enhancement, strong anatomical priors in the degraded input render dense noise schedules largely redundant. Leveraging this insight, we propose a sparse voxel-space diffusion framework that trains and samples on a compact set of uniformly subsampled timesteps. The network predicts clean data directly on the data manifold, supervised in velocity space for stable gradient scaling. A lightweight Structure-aware Trajectory Modulation (STM) module recalibrates time embeddings at each network block based on local anatomical content, enabling structure-adaptive denoising over the shared sparse schedule. Operating directly in voxel space, our framework preserves fine anatomical detail without lossy compression while achieving up to $10\times$ training acceleration. Experiments on four datasets spanning CT, PET, and MRI demonstrate state-of-the-art performance on both denoising and super-resolution tasks. Our code is publicly available at: https://github.com/mirthAI/sparse-3d-diffusion.

cs.CV↗

arXiv2026

PARM: Pipeline-Adapted Reward Model

Reward models (RMs) are central to aligning large language models (LLMs) with human preferences, powering RLHF and advanced decoding strategies. While most prior work focuses on single-step generation, real-world applications increasingly adopt multi-stage LLM pipelines, where effective reward guidance remains underexplored. We investigate this through code generation for combinatorial optimization, constructing a pipeline that integrates reward models into both formulation and solution stages. We identify a critical challenge: inconsistency between reward model predictions and actual pipeline execution outcomes. To address this, we propose the Pipeline-Adapted Reward Model (PARM), which leverages pipeline-specific data and direct preference optimization to align rewards with downstream feedback. We instantiate PARM as a two-stage pipeline (formulation -> code generation) and evaluate it on four public optimization benchmarks, measuring execution rate and solving accuracy against baselines and sampling methods. A supplementary cross-domain experiment on GSM8K assesses transferability. Results demonstrate that PARM consistently improves pipeline output quality and stability, providing new insights into reward modeling for multi-stage LLM reasoning.

cs.AI↗

arXiv2026

Kumo: A Security-Focused Serverless Cloud Simulator

Serverless computing abstracts infrastructure management but also obscures system-level behaviors that can introduce security risks. Prior work has shown that serverless platforms are vulnerable to attacks exploiting shared execution environments, including attacker--victim co-location and denial-of-service through resource contention, yet analyzing these risks on production platforms is difficult due to limited observability, high cost, and lack of experimental control, while existing simulators primarily focus on performance and cost rather than security. We present Kumo, a security-focused simulator for serverless platforms that enables controlled, reproducible analysis of security risks arising from scheduling and resource sharing decisions. Kumo models invocation arrivals, scheduler placement, container reuse, resource contention, and queuing within a discrete-event framework, explicitly representing attackers and victims as first-class entities and providing metrics such as co-location probability, time to first co-location, invocation drop rate, and tail latency. Through two case studies, we show that scheduler choice is a first-order factor for co-location attacks, inducing orders-of-magnitude differences under identical workloads, while Denial-of-Service behavior is largely governed by system-level factors such as service time, queuing policy, and cluster capacity once contention dominates. These results highlight the need to distinguish scheduler-driven isolation risks from broader resource exhaustion vulnerabilities and position Kumo as a flexible foundation for systematic, security-aware exploration of serverless platforms.

cs.CR↗

arXiv2026

How Well Does Generative Recommendation Generalize?

A widely held hypothesis for why generative recommendation (GR) models outperform conventional item ID-based models is that they generalize better. However, there is few systematic way to verify this hypothesis beyond a superficial comparison of overall performance. To address this gap, we categorize each data instance based on the specific capability required for a correct prediction: either memorization (reusing item transition patterns observed during training) or generalization (composing known patterns to predict unseen item transitions). Extensive experiments show that GR models perform better on instances that require generalization, whereas item ID-based models perform better when memorization is more important. To explain this divergence, we shift the analysis from the item level to the token level and show that what appears to be item-level generalization often reduces to token-level memorization for GR models. Finally, we show that the two paradigms are complementary. We propose a simple memorization-aware indicator that adaptively combines them on a per-instance basis, leading to improved overall recommendation performance.

cs.IR↗

arXiv2026

Towards Privacy-Preserving Machine Translation at the Inference Stage: A New Task and Benchmark

Current online translation services require sending user text to cloud servers, posing a risk of privacy leakage when the text contains sensitive information. This risk hinders the application of online translation services in privacy-sensitive scenarios. One way to mitigate this risk for online translation services is introducing privacy protection mechanisms targeting the inference stage of translation models. However, compared to subfields of NLP like text classification and summarization, the machine translation research community has limited exploration of privacy protection during the inference stage. There is no clearly defined privacy protection task for the inference stage, dedicated evaluation datasets and metrics, and reference benchmark methods. The absence of these elements has seriously constrained researchers' in-depth exploration of this direction. To bridge this gap, this paper proposes a novel "Privacy-Preserving Machine Translation" (PPMT) task, aiming to protect the private information in text during the model inference stage. For this task, we constructed three benchmark test datasets, designed corresponding evaluation metrics, and proposed a series of benchmark methods as a starting point for this task. The definition of privacy is complex and diverse. Considering that named entities often contain a large amount of personal privacy and commercial secrets, we have focused our research on protecting only the named entity's privacy in the text. We expect this research work will provide a new perspective and a solid foundation for the privacy protection problem in machine translation.

cs.CL↗

arXiv2026

Advancing Cancer Prognosis with Hierarchical Fusion of Genomic, Proteomic and Pathology Imaging Data from a Systems Biology Perspective

To enhance the precision of cancer prognosis, recent research has increasingly focused on multimodal survival methods by integrating genomic data and histology images. However, current approaches overlook the fact that the proteome serves as an intermediate layer bridging genomic alterations and histopathological features while providing complementary biological information essential for survival prediction. This biological reality exposes another architectural limitation: existing integrative analysis studies fuse these heterogeneous data sources in a flat manner that fails to capture their inherent biological hierarchy. To address these limitations, we propose HFGPI, a hierarchical fusion framework that models the biological progression from genes to proteins to histology images from a systems biology perspective. Specifically, we introduce Molecular Tokenizer, a molecular encoding strategy that integrates identity embeddings with expression profiles to construct biologically informed representations for genes and proteins. We then develop Gene-Regulated Protein Fusion (GRPF), which employs graph-aware cross-attention with structure-preserving alignment to explicitly model gene-protein regulatory relationships and generate gene-regulated protein representations. Additionally, we propose Protein-Guided Hypergraph Learning (PGHL), which establishes associations between proteins and image patches, leveraging hypergraph convolution to capture higher-order protein-morphology relationships. The final features are progressively fused across hierarchical layers to achieve precise survival outcome prediction. Extensive experiments on five benchmark datasets demonstrate the superiority of HFGPI over state-of-the-art methods.

cs.CV↗