Watermark radioactivity testing type of methods can detect whether a model was trained on watermarked documents, and have become key tools for protecting data ownership in the fine-tuning of large language models (LLMs). Existing works have proved their effectiveness in centralized LLM fine-tuning. However, this type of method faces several challenges and remains underexplored in federated learning (FL), a widely-applied paradigm for fine-tuning LLMs collaboratively on private data across different users. FL mainly ensures privacy through secure aggregation (SA), which allows the server to aggregate updates while keeping clients' updates private. This mechanism preserves privacy but makes it difficult to identify which client trained on watermarked documents. In this work, we propose FedAttr, a new client-level attribution protocol for FL. FedAttr identifies which clients trained on watermarked data via a paired-subset-difference mechanism, while preserving the privacy guarantees of SA and FL performance. FedAttr proceeds in three steps: (i) estimate each client's update by differencing two SA queries, (ii) score the estimate with the watermark detector via differential scoring, and (iii) combine scores across rounds via Stouffer method. We theoretically show that FedAttr produces an unbiased estimator of each client's update with bounded mutual information leakage (i.e., $O(d^*/N)$ per-round update). Moreover, FedAttr empirically achieves 100% TPR and 0% FPR, outperforming all baselines by at least 44.4% in TPR or 19.1% in FPR, with only 6.3% overhead relative to FL training time. Ablation studies confirm that FedAttr is robust to protocol parameters and configurations.

While real-world applications of reinforcement learning (RL) are becoming increasingly popular, the security of RL systems deserve more attention and exploration. In particular, recent work has revealed that RL agents are vulnerable to backdoor attacks, where a victim agent behaves normally under standard conditions but executes malicious actions when a specific trigger is activated. Existing backdoor defenses for RL either require access to the agent’s internal parameters, operate only at the model or trajectory level, or are limited to specific attack types. To ensure the security of RL agents, we propose PolicyGuard, a test-time step-level backdoor defense which leverages Gaussian Process (GP) posterior variance and adapts pseudo trajectories to enable uncertainty computation for individual time step. Besides, we also provide theoretical foundations to explain the efficacy of GP posterior variance. Extensive experiments across seven RL games demonstrate that PolicyGuard achieves state-of-the-art detection performance in most cases, with average AUROC of 0.856 for perturbation-based attacks and 0.859 for adversary-agent attacks.

Graph Neural Networks (GNNs) serve as the backbone for high-stakes applications in Machine-Learning-as-a-Service (MLaaS). Still, their black-box deployment exposes them to Model Extraction (ME) attacks, in which adversaries steal intellectual property by querying APIs. Existing defenses suffer from a critical ”Euclidean bias”: they transfer image-based strategies (e.g., random noise) to graphs, ignoring the complex topological dependencies between nodes, which often results in severe utility degradation. Passive methods like watermarking also fail to prevent theft in real time. To bridge this gap, we propose GraphRP (Graph Reprogramming Protection), a proactive defense framework that repurposes Model Reprogramming for security. Unlike static perturbations, GraphRP introduces a Structure-Aware Gating Mechanism driven by learnable topological prototypes. This creates a dynamic ”structural firewall” that selectively modulates the model's decision boundary: it preserves fidelity for benign queries residing on the training manifold, while maximizing the Fisher Information along the perturbation direction for adversarial queries. Under standard assumptions (bounded loss, optimal attacker, and local second-order approximation), we prove a lower bound on the attacker's estimation error that increases with the structural sensitivity of the reprogramming noise. Extensive experiments on both hard-label and soft-label ME attacks demonstrate that GraphRP significantly degrades attack effectiveness while preserving benign utility.

Lithium iron phosphate electrodes are widely used in lithium-ion batteries, but their low intrinsic electronic conductivity can produce non-uniform reaction and heat-generation distributions during high-rate operation. In this study, a deterministic three-dimensional microscale electrochemical model with one-way thermal coupling was developed to examine how electronic conduction, ionic transport, active-particle size, temperature, and conductive-binder-domain (CBD) distribution affect local electrode behavior. The model reproduces the rate-dependent voltage response of LFP half-cells and provides spatial information that is not directly available from macroscopic discharge curves. Under 3 C discharge, the solid-phase potential drop reached 198.75 mV at 0.6 depth of discharge, whereas the liquid-phase potential drop was only 4.13 mV, indicating that solid-phase electronic conduction strongly affected reaction localization. The CBD network accounted for 87.17–91.22% of the total Joule heat over the examined temperature range. Reducing particle size improved macroscopic capacity and increased the local current-density peak near the current collector, while decreasing the through-thickness variation of reaction current. The CBD gradient results further showed that abrupt low-conductivity regions can introduce local polarization losses under fixed total CBD content. These results provide a controlled microscale analysis of coupled transport and heat-generation behavior in LFP electrodes and support the design of conductive networks with continuous electron-transport pathways.

Continual learning (CL) aims to acquire new knowledge from a non-stationary data stream while retaining performance on previously learned tasks. Memory-based replay methods mitigate catastrophic forgetting by storing and revisiting past samples, but their effectiveness is fundamentally constrained by limited memory capacity, as each stored example represents only a single data instance. In this work, we propose data reassembly for CL, a new paradigm that significantly increases memory efficiency by reassembling composite replay samples from existing training data. Instead of storing raw training examples, we partition the current task training data into elementary patches and dynamically reassemble them into coherent replay instances through an energy-based optimization framework. The proposed objective jointly enforces semantic compatibility with target labels and global consistency among assembled patches. To make this optimization tractable, we derive an efficient variational inference algorithm that constructs a compact yet diverse set of reassembled samples for replay. Extensive theoretical analysis and experiments across multiple CL benchmarks demonstrate that data reassembly consistently outperforms existing memory-based approaches, achieving stronger retention of past knowledge while maintaining competitive computational efficiency.

Test-time scaling (TTS) has become an effective approach for improving large language model performance by allocating additional computation during inference. However, existing TTS strategies are largely hand-crafted: researchers manually design reasoning patterns and tune heuristics by intuition, leaving much of the computation-allocation space unexplored. We propose an environment-driven framework, AutoTTS, that changes what researchers design: from individual TTS heuristics to environments where TTS strategies can be discovered automatically. The key to AutoTTS lies in environment construction: the discovery environment must make the control space tractable and provide cheap, frequent feedback for TTS search. As a concrete instantiation, we formulate width--depth TTS as controller synthesis over pre-collected reasoning trajectories and probe signals, where controllers decide when to branch, continue, probe, prune, or stop and can be evaluated cheaply without repeated LLM calls. We further introduce beta parameterization to make the search tractable and fine-grained execution trace feedback to improve discovery efficiency by helping the agent diagnose why a TTS program fails. Experiments on mathematical reasoning benchmarks show that the discovered strategies improve the overall accuracy--cost tradeoff over strong manually designed baselines. The discovered strategies generalize to held-out benchmarks and model scales, while the entire discovery costs only $39.9 and 160 minutes. Our data, and code will be open-source at https://github.com/zhengkid/AutoTTS.

While collaborative filtering delivers predictive accuracy and efficiency, and Large Language Models (LLMs) enable expressive and generalizable reasoning, modern recommendation systems must bring these strengths together. Growing user expectations, such as natural-language queries and transparent explanations, further highlight the need for a unified approach. However, doing so is nontrivial. Collaborative signals are often token-efficient but semantically opaque, while LLMs are semantically rich but struggle to model implicit user preferences when trained only on textual inputs. This paper introduces Item-ID + Natural-language Mixture-of-Experts Language Model (IDIOMoE), which treats item interaction histories as a native dialect within the language space, enabling collaborative signals to be understood in the same way as natural language. By splitting the Feed Forward Network of each block of a pretrained LLM into a separate text expert and an item expert with token-type gating, our method avoids destructive interference between text and catalog modalities. IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets, while preserving the text understanding of the pretrained model.

Proton exchange membrane fuel cells (PEMFCs) are highly susceptible to performance limitations at high current densities due to competing gas and liquid transport mechanisms. While microporous layers (MPLs) reduce contact resistance, they often exacerbate the risk of flooding. This study leverages the topological advantages of highly interconnected cage-like MPLs. By integrating X-ray computed tomography (X-CT) with digital techniques (reconstructed geometry and numerical simulation), it reveals the mass transfer mechanisms within cross-scale composite microstructures. Considering in-situ compression and surface wettability, the research delves into the influence of MPL thickness and crack morphology on multiphase flow dynamics. The results indicate that despite the significant non-steady-state velocity fluctuations and reduced liquid permeability caused by cascading Haines jumps due to increased MPL thickness, its unique high-porosity cage-like network structure still maintains robust oxygen diffusion performance. In-situ compaction further weakens the macropore transport capacity of the gas diffusion layer (GDL), leading to a significant reduction in permeability. In contrast, fractures effectively break through transport bottlenecks. Vertical through-holes maximize oxygen diffusion efficiency through the shortest path effect, while 45 degrees inclined fractures optimize gas-liquid convective permeability while enabling directed high-speed drainage of liquid water. This study elucidates the regulation mechanism of microstructural parameters across scales, providing theoretical guidance for designing next-generation gas diffusion layers that combine high gas-phase diffusion with efficient drainage capabilities.

Accurate in vivo prediction of neuropathology is critical for advancing diagnosis and treatment of Alzheimer’s disease and related dementias (ADRDs). As many individuals with ADRDs have mixed pathologies (β-amyloid, pathologic tau, cerebrovascular disease, vascular brain injury, pathologic TDP-43, hippocampal sclerosis, Lewy bodies), there is interest in determining how accurately we can infer these pathologic changes from clinical data, biofluid assays (e.g., CSF), and neuroimaging. Here we evaluated automated machine learning models trained on data curated by the AD Sequencing Project Phenotype Harmonization Consortium (N=7,894 individuals), to predict 26 autopsy-confirmed neuropathological outcomes. Predictors included in vivo clinical and cognitive composite scores, brain measures from 3D structural MRI and diffusion tensor imaging, image-derived measures of white matter hyperintensities (WMH), and CSF biomarkers. Predictive models were trained using ensemble learning with stratified cross-validation. We assessed performance using Spearman’s rank correlation and Matthews correlation coefficient, to accommodate co-occurring pathologic changes. The added value of neuroimaging and CSF versus clinical features alone was quantified. Braak stage was among the most consistently predicted outcomes. CSF biomarkers best predicted β-amyloid and tau pathology, but diffusion MRI metrics best captured vascular brain injury and white matter injury, and outperformed clinical and cognitive measures and anatomical MRI in predicting Lewy body disease. Anatomical measures from structural MRI outperformed standard clinical assessments in assessing neurodegeneration and hippocampal sclerosis, and WMH complemented cognitive measures in predicting TDP-43 pathology. These results establish a baseline for comparing modalities for inferring neuropathology.

The gas diffusion layer (GDL) of proton exchange membrane fuel cells (PEMFCs) is a critical component for the transport of reactants. The efficiency of reactant gas transport remains a major technical challenge in the field today. The anisotropic structure of the GDL gives rise to substantial variations of gas diffusion as well as permeability in different directions. The study employs X-CT technology to obtain the actual GDL's geometry, aiming to investigate a spatial structure at the microscale and its gas transport characteristics. The computational fluid dynamics (CFD) method is used to simulate and study the gas diffusivity and gas permeability of GDL with four different thicknesses. The numerical simulation results show that the diffusivity and permeability in the through-plane (TP) direction are lower than those in the in-plane (IP) direction. Moreover, the effective diffusion coefficient (EDC) decreases with increasing thickness, but is also dependent on the solid fibre structure of GDL. Horizontal alignment of the carbon fibers and the disc-shaped adhesive contributes to the anisotropy between the TP and IP directions, resulting in anisotropic gas transport. The purpose of the study is to supply critical references for manufacturing techniques and optimization of gas transport in GDLs.

Predicting the trajectory of clinical decline in aging individuals is a pressing challenge, especially for people with mild cognitive impairment, Alzheimer’s disease, Parkinson’s disease, or vascular dementia. Accurate predictions can guide treatment decisions, identify risk factors, and optimize clinical trials. In this study, we compared two deep learning approaches for forecasting changes, over a 2-year interval, in the Clinical Dementia Rating scale ‘sum of boxes’ score (sobCDR), as a continuous outcome (regression). This is a key metric in dementia research and clinical trials, and scores range from 0 (no impairment) to 18 (severe impairment). To predict decline, we trained a hybrid convolutional neural network (CNN) that integrates 3D T1-weighted brain MRI scans with tabular clinical and demographic features (including age, sex, body mass index (BMI), and baseline sobCDR). We benchmarked its performance against AutoGluon, an automated multimodal machine learning framework that selects an appropriate neural network architecture (an ‘autoML’ approach). We evaluated the models using data from 2,319 unique participants drawn from three independent cohorts—ADNI, OASIS-3, and NACC. For each participant, we used one T1-weighted brain MRI scan along with corresponding clinical and demographic information. Our results demonstrate the importance of combining image and tabular data in predictive modeling for this clinical application. Deep learning algorithms can fuse information from image-based brain signatures and tabular clinical data, with potential for personalized prognostics in aging and dementia. Rather than concluding that multimodal fusion uniformly improves performance, our results show that deep learning applied to volumetric MRI data may struggle to add predictive value, particularly when clinical covariates explain substantial variance and provide a strong baseline. In other conditions and tasks, it may help to have a hybrid system that can learn from both data types, and their relative value may be different. Conversely, AutoML-based multimodal fusion provides a robust baseline when tabular data already provide strong predictive value for the task. These insights clarify how different multimodal strategies could be selected in clinical prognostic applications.

IntroductionEarly detection of Alzheimer's disease (AD) requires models that combine brain structure changes with genetic risk, but existing methods struggle to align these different data types.MethodsWe present R-GenIMA, an interpretable multimodal large language model that pairs a region-of-interest vision transformer with genetic prompting to jointly analyze structural MRI and single nucleotide polymorphisms (SNPs). Each brain region becomes a visual token and SNP profiles are encoded as structured text, letting the model link regional atrophy to genetic factors through cross-modal attention. Tested on the ADNI cohort, R-GenIMA performs well in classifying four groups: normal cognition, subjective memory concerns, mild cognitive impairment, and AD.ResultsBeyond accuracy, it produces biologically meaningful explanations, identifying stage-specific brain regions and genes. The model consistently highlighted known AD risk genes (APOE, BIN1, CLU, RBFOX1) and revealed stage-specific patterns: striatal involvement in subjective decline, frontotemporal changes in early impairment, and broad network disruption in AD.DiscussionThese results show that interpretable multimodal AI can integrate imaging and genetics to reveal disease mechanisms, providing a foundation for clinical tools that enable earlier risk assessment and inform precision treatment in Alzheimer's disease.

We introduce VGA-BenchV2, an extended human-aligned benchmark and optimization framework for jointly evaluating and improving video generation quality and aesthetic value. Built upon VGA-Bench, VGA-BenchV2 preserves the original fine-grained taxonomy with two primary dimensions-Aesthetic and Generation-and 52 sub-dimensions. Guided by this taxonomy, we curate 1,016 diverse prompts and collect over 60,000 videos generated by 12 mainstream video generation models. More importantly, VGA-BenchV2 substantially expands human-labeled supervision by adding 36,000 task-level annotations, including 16,200 for aesthetic quality, 13,200 for aesthetic tagging, and 6,600 for generation quality, corresponding to 13.46x, 11.15x, and 1.55x scale-ups over VGA-Bench, respectively. Leveraging this enlarged annotation corpus, we develop a hybrid evaluator architecture consisting of VAQA-Net for continuous aesthetic scoring and two Qwen-based Large Vision-Language Model evaluators, VTag-Net and VGQA-Net, for aesthetic tagging and generation quality assessment. Extensive experiments demonstrate strong alignment with human judgments across diverse generation models. Beyond evaluation, VGA-BenchV2 further introduces an evaluation-to-optimization pipeline, where the learned aesthetic evaluator serves as a reward model for reinforcement learning-based generator fine-tuning. This closes the loop from benchmark construction and human supervision to automated evaluation and model optimization, enabling video generators to improve not only in realism but also in aesthetic quality and human preference alignment. Resources are available at https://huggingface.co/datasets/BestiVictoryLab/VGA-Bench.

Conventional multi-stage cell tracking approaches rely heavily on detection or segmentation in each frame as a prerequisite, requiring substantial resources for high-quality segmentation masks and increasing the overall prediction time. To address these limitations, we propose CAP, a novel end-to-end one-stage framework that reimagines cell tracking by treating Cell as Point. Unlike traditional methods, CAP eliminates the need for explicit detection or segmentation, instead jointly tracking cells for sequences in one stage by leveraging the inherent correlations among their trajectories. This simplification reduces both labeling requirements and pipeline complexity. However, directly processing the entire sequence in one stage poses challenges related to data imbalance in capturing cell division events and long sequence inference. To solve these challenges, CAP introduces two key innovations: (1) adaptive event-guided (AEG) sampling, which prioritizes cell division events to mitigate the occurrence imbalance of cell events, and (2) the rolling-as-window (RAW) inference strategy, which ensures continuous and stable tracking of newly emerging cells over extended sequences. By removing the dependency on segmentation-based preprocessing while addressing the challenges of imbalanced occurrence of cell events and long-sequence tracking, CAP demonstrates promising cell tracking performance and is 8 to 32 times more efficient than existing methods. The code and model checkpoints are available at https://github.com/YXSong000/CAP.

Current whole-genome spatial transcriptomics (ST) platforms, which range from multicellular (e.g., Visium) to subcellular resolution (e.g., Visium HD, Stereo-seq), face significant challenges in accurately decomposing or aggregating spots into single cells-the fundamental units of biological activity. To date, most existing computational methods operate at the spot level, and no method has effectively reconstructed transcriptomes at single-cell resolution. In this study, we present STARS (Spatial Transcriptomics across Resolutions for Single Cells). Leveraging Vision Transformer model and contrastive learning, STARS combines high-resolution histology images with spot-level transcriptomics data to reconstruct single-cell-level gene expression from multicellular and subcellular platforms. We demonstrate the advantage of our single-cell-level method using in-house datasets of mouse lung from 3 ST platforms (Visium, Visium HD and Stereo-seq) and additional public datasets. STARS is applied at tissue, individual cell, and molecular levels. At the tissue level, STARS identifies regions of interest, such as specific tissue structures or immune regions. At the individual cell level, STARS identifies immune cells-including CD4/CD8 T cells and cancer-associated fibroblasts (CAFs)-distinguishing two CAF subtypes and macrophage subtypes, including the rare SPP1+ macrophages. Additionally, STARS identifies tertiary lymphoid structures in colorectal cancer, linked to improved clinical outcomes, and detects shifts in immune cell populations between influenza infection and bacterial superinfection in mouse lung tissue. At the molecular level, we demonstrate improved separation of cell types, and differential gene expression analysis identified cell-type-specific markers with greater accuracy. These results underscore the robustness of our method, providing biologically relevant insights into tissue architecture and gene expression at the single-cell level, advancing the downstream analysis of spatial transcriptomics across resolutions.

Alzheimer's disease (AD) progresses as a continuous biological process, whereas most existing neuroimaging-based artificial intelligence methods remain limited to discrete diagnosis or clinical score prediction from cross-sectional imaging. In this work, we propose Disease Continuum Positioning (DCP), a longitudinal Bayesian Learning framework that continuously estimates disease severity from longitudinal diffusion tensor imaging (DTI). Specifically, DCP models disease severity as a low-dimensional probabilistic latent variable by jointly integrating longitudinal observations with weak clinical supervision, from which the proposed Disease Continuum Score (DCS) is derived to quantify an individual's position along the Alzheimer's disease continuum together with its associated uncertainty. Extensive experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort demonstrate that DCP consistently outperforms representative disease progression methods. More importantly, comprehensive validation analyses show that DCS accurately characterizes disease severity, exhibits strong clinical relevance, preserves longitudinal disease evolution, and predicts future disease conversion. These results suggest that DCS provides a quantitative imaging-derived representation for continuous assessment of Alzheimer's disease progression beyond conventional diagnostic labels and clinical scores.

The performance of solid oxide electrolysis cells (SOECs) is closely related to its flow channel structure, stoichiometric ratio, and operating temperature. In this study, multiphysics numerical models were developed, incorporating coupled heat and mass transport alongside electrochemical reaction processes. The model validity was confirmed by comparing simulated results with experimentally measured I-V curves. The impacts of the fuel-to-air stoichiometric ratio (F:A from 2:1 to 2:4), channel aspect ratio (L:W from 1:1 to 4:1), temperature (from 873.15 K to 1073.15 K), and flow arrangements (co-current vs. counter-current) on the performance of a single-channel electrolyzer were systematically investigated, complemented by an analysis of multi-channel behavior under cross-flow conditions. The findings reveal that among the investigated parameters, temperature exerts the most significant influence on cell performance. Increasing the temperature facilitates the substitution of electrical energy with thermal energy, reducing the cell voltage from 1.5364 V to 1.1142 V at 1.2 A/cm2. Furthermore, varying the stoichiometric ratio effectively improves the oxygen partial pressure in the catalyst layer, thereby reducing concentration polarization. At the same current density, the required cell voltage decreased from 1.5364 V to 1.5153 V. Increasing the channel aspect ratio improves mass transport, reducing the required cell voltage from 1.5444 V to 1.5053 V. In contrast, flow arrangements were found to have a negligible impact on overall performance, though the counter-flow arrangement demonstrated marginal superiority over the co-flow arrangement.

Highlights• Introduces SPBA: the first patch-wise backdoor for 3D point clouds.• Employs curvature-based scoring to select imperceptible complex patches.• Applies patch-wise spectral triggers (GFT) to minimize distortion and cost.• Demonstrates superior trade-off in effectiveness, stealthiness, and robustness.Stealthy Patch-Wise Backdoor Attack in 3D Point Cloud via Curvature AwarenessBackdoor attacks pose a severe threat to deep neural networks (DNNs) by implanting hidden backdoors that can be activated with predefined triggers to manipulate model behaviors maliciously. Existing 3D point cloud backdoor attacks primarily rely on sample-wise global modifications, which suffer from low imperceptibility. Although optimization can improve stealthiness, optimizing sample-wise triggers significantly increases computational cost. To address these limitations, we propose the Stealthy Patch-Wise Backdoor Attack (SPBA), the first patch-wise backdoor attack framework for 3D point clouds. Specifically, SPBA decomposes point clouds into local patches and employs a curvature-based imperceptibility score to guide trigger injection into visually less sensitive patches. By optimizing a unified patch-wise trigger that perturbs spectral features of selected patches, SPBA significantly enhances optimization efficiency while maintaining high stealthiness. Extensive experiments on ModelNet40 and ShapeNetPart further demonstrate that SPBA surpasses prior state-of-the-art backdoor attacks in both attack effectiveness and resistance to defense methods. The code is available at https://github.com/HazardFY/SPBA.

Large multimodal models (LMMs) are increasingly adopted as judges in multimodal evaluation systems due to their strong instruction following and consistency with human preferences. However, their ability to follow diverse, fine-grained evaluation criteria remains underexplored.We develop Multi-Crit, a benchmark for evaluating multimodal judges on their capacity to follow pluralistic criteria and produce reliable criterion-level judgments. Covering both open-ended generation and verifiable reasoning tasks, Multi-Crit is built through a rigorous data curation pipeline that gathers challenging response pairs with multi-criterion human annotations. It further introduces three novel metrics for systematically assessing pluralistic adherence, criterion-switching flexibility, and the ability to recognize criterion-level preference conflicts.Comprehensive analysis of 25 LMMs reveals that 1) proprietary models still struggle to maintain consistent adherence to pluralistic criteria—especially in open-ended evaluation; 2) open-source models lag further behind in flexibly following diverse criteria; and 3) critic fine-tuning with holistic judgment signals enhances visual grounding but fails to generalize to pluralistic criterion-level judgment.Additional analyses on reasoning fine-tuning, test-time scaling, and boundary consistency between open-source and proprietary models further probe the limits of current multimodal judges.As a pioneering study, Multi-Crit lays the foundation for building reliable and steerable multimodal AI evaluation.