The Pearl River Delta region is densely dotted with sand dune sites，whose subsistence economies hold significant importance for understanding the maritime adaptation and civilizational evolution of ancient populations in South China. However，direct evidence of marine resource utilization by ancient populations has long been scarce due to the poor preservation of faunal and floral remains. Lipid analysis of pottery effectively traces marine processing，yet this method has not been applied to such sites. In this study，we conducted lipid analysis for the first time on pottery shards (among them, 2 samples exhibited visible residues, and 58 samples showed adsorbed residues.) excavated from the Taishan XinCun site to reveal food processing practices and subsistence strategies. The GC⁃MS results indicate that residues exhibit characteristics of mixed animal and plant origins，with markers potentially derived from cereal seed oils detected in some samples. The GC⁃C⁃IRMS results revealed that lipids in approximately 87% of samples originated from marine animals (fish，shellfish，etc.)，while a small number of samples derived from terrestrial non⁃ruminant animal fat. Combining lipid composition with pottery characteristics，pottery Guan and Fu likely served as cooking pottery for processing mixed marine and terrestrial resources，while other pottery，bowls etc. as serving pottery. Thus，around 5000 years ago，Xincun site had developed a mixed subsistence strategy dominated by marine fishing and hunting，supplemented by terrestrial hunting，plant gathering. This reflects a high degree of adaptation to the coastal environment and laid the foundation for the formation of early maritime culture in the region.

Elevated CO₂ levels (eCO₂) and heavy metal contamination are two environmental factors that significantly affect soil carbon cycle.However,their combined effects on the fate and transformation of plant⁃derived organic matter in soil remain unclear. Using ¹⁴C⁃labelled wheat roots as a model for plant⁃derived organic matter,we investigated the fate and transformation of plant⁃derived organic matter in Cu/Cd contaminated soils under eCO₂ conditions.In the 180⁃day study,eCO₂ significantly increased the concentration of DTPA⁃Cu,while having a limited effect on Cd,suggesting that eCO₂ may enhance the bioavailability and potential risk of certain heavy metals. Compared with the control group,Cu/Cd treatments inhibited the mineralization of plant⁃derived organic matter during early incubation; however,owing to the intrinsic resilience of the soil,this inhibitory subsidedby the middle and late stages. Furthermore,eCO₂ significantly accelerated the degradation and transformation of labile soil components,such as fulvic acid⁃like substances and tyrosine,thereby affecting the stability of the soil carbon pool. Concurrently,eCO₂ increased the proportion of humin in the transformation products,which may accelerated the humification of plant⁃derived organic matter. In contrast,Cu hindered the formation of stable humin from plant‑derived organic matter,potentially suppressing the humification process. These findings provide new insights into soil carbon transformation dynamics under the combined stress of eCO₂ and Cd/Cu contamination,providing a scientific basis for the ecological risk assessment of heavy metal⁃contaminated soils and for the development of carbon management strategies under eCO₂ condition.

Head⁃related transfer function (HRTF) databases are important data resources for spatial audio rendering and personalized HRTF modeling. This study compared directional and subject⁃level clustering structures in the SONICOM，PHRTF⁃NJU，and HUTUBS databases using consistent feature representations and similarity definitions across the three databases. For directional clustering，HRTF magnitude spectra smoothed on the equivalent rectangular bandwidth (ERB) scale were used as features，and k⁃means clustering was applied to analyze directional spectral patterns. For subject⁃level clustering，an inter⁃subject similarity matrix was constructed using the mean normalized cross⁃correlation coefficient (MNCC) computed over both ears and all directions，followed by average⁃linkage hierarchical clustering to analyze overall inter⁃subject similarity. Under the feature set and evaluation metrics used in this study，the three databases showed similar trends as the number of directional clusters varied: a two⁃cluster solution yielded a relatively stable coarse⁃grained structure，whereas a three⁃cluster solution provided a finer spatial partition. Subject⁃level clustering in all three databases exhibited a common threshold⁃dependent transition from fragmentation to progressive merging. Overall，HRTFs from the three databases exhibited similar coarse⁃grained directional structures while retaining database⁃specific differences in similarity scale and local structure.

Functional ultrasound imaging (fUS)，based on ultrafast Doppler imaging and the neurovascular coupling mechanism，can record dynamic changes in cerebral blood volume (CBV) with high spatiotemporal resolution over a large field of view，providing a novel imaging approach for studying the functions of deep brain regions in rodents. To investigate activity in the subcortical visual pathway of rats evoked by visual stimulation，we employed a plane‑wave‑based composite ultrafast ultrasound imaging system to perform synchronized imaging during monocular blue‑light flicker stimulation in anesthetized adult male Sprague‑Dawley rats. The visual stimulus frequency was 12 Hz，with each cycle comprising 5 s of stimulation followed by 55 s of rest. We acquired data from ten stimulus cycles in the coronal planes containing the superior colliculus (SC) and the lateral geniculate nucleus (LGN)，and extracted stimulus‑related CBV changes using singular value decomposition for clutter suppression，power Doppler computation，and temporal averaging. Results showed that stimulation of the left eye primarily induced increases in CBV in the right SC and LGN，whereas stimulation of the right eye predominantly elicited responses in the left SC，suggesting a clear contralateral dominance in rat visual information processing. The amplitude of the SC response was approximately 5%~10%，with a peak around 9%，and recovery after stimulation was relatively slow. In contrast，the LGN exhibited a peak of about 6%，with a response more closely synchronized with the stimulus input and returning to baseline within approximately 5 s post‑stimulation. These findings demonstrate that fUS can noninvasively visualize stimulus‑evoked responses and their spatial distribution in deep visual‑related structures of rats，providing experimental evidence and methodological guidance for future studies of the functional hemodynamics of visual pathways in rodents.

Addressing challenges such as the high proportion of small targets，severe target overlap，complex background interference，and random distribution of target orientations in security⁃check X⁃ray images，we propose a directional object detection method based on a dual enhancement mechanism⁃combining semantic and sample augmentation. In terms of semantic enhancement，we adopt the You Only Look Once⁃Oriented Bounding Box (YOLO⁃OBB) framework，which is designed for rotated bounding box detection，as the backbone model. We further introduce a frozen Image Encoder from the Segment Anything Model (SAM) as a semantic auxiliary branch and employ a channel⁃attention fusion module to achieve adaptive fusion of heterogeneous features，thereby enhancing the model's ability to capture both target structural information and high⁃level semantic cues. For sample augmentation，we design a Context⁃Constrained Copy⁃Paste Augmentation (CCPA) strategy that performs controlled augmentation of small⁃target instances under constraints on spatial occupancy，collision detection，and Poisson blending. This strategy expands the distribution of training samples across scale，location，and orientation spaces，alleviating the issues of insufficient small⁃target samples and long⁃tail distribution. Additionally，we introduce a Gaussian⁃smoothed focal loss to improve the stability of rotated bounding box regression. Experimental results demonstrate that our proposed method outperforms the baseline model in directional object detection tasks within security⁃check scenarios，and ablation studies confirm the effectiveness and complementarity of each module.

Retinal vessel segmentation is a fundamental task in fundus image analysis，playing a vital role in computer⁃aided diagnosis and early screening for fundus disea ses. However，due to the slender，continuous structure，dense branching，and blurred boundaries of blood vessels，existing models⁃ranging from conventional convolutional neural net works to Transformer⁃based approaches⁃tend to suffer from missed detection and fragmentation，especially in regions with tiny vessels and ambiguous boundaries. This is because thes e methods lack sufficient capability to explicitly model such uncertain and fuzzy areas. To address these challenges，this paper proposes a fuzzy⁃guided dual⁃branch Transformer⁃UNet framework for retinal vessel segmentation. The core idea of our approach is to explicitly model fuzzy regions and integrate this concept into both the structural design and optim ization constraints throughout the framework. First，we construct a fuzzy membership map from the input fundus images without requiring additional annotations. This map characterizes regions with unstable local textures and blurry boundaries，providing a consistent prior that can be applied consistently across both the training and inference stages. Second，while preserving the global representation capability of the Transformer encoder，we introduce a standard branch and a fuzzy branch for parallel decoding. A fuzzy gating mechanism is employed to modulate multi⁃scale skip connections，enabling the fuzzy branch to focus on the fine reconstruction of micro⁃vessels and weak boundaries. Next，we design a pixel⁃wise adaptive fusion module that dynamically integrates the outputs of the two branches，achieving complementary collaboration between structural restoration and boundary refinement. Finally，we incorporate region⁃weighted supervision for fuzzy areas and inter⁃branch consistency constraints into the loss function. These additions enhance the discriminative power of uncertain regions from an optimization perspective and help suppress boundary jitter. In addition，we conduct a large⁃scale evaluation on the FIVES dataset，which contains 800 high⁃resolution fundus images from four different disease categories. We further perform six progressive ablation configurations and a frozen⁃weight one⁃factor⁃at⁃a⁃time sensitivity analysis to examine the contributions of the dual⁃branch decoder，fuzzy gating，adaptive fusion，fuzzy⁃aware losses，and prior⁃parameter settings. Experimental results demonstrate that the proposed method maintains stable segmentation performance across various data distributions and achieves favorable results in metrics such as accuracy and F1⁃score. In particular，it exhibits stronger robustness in segmenting vessel boundaries that are blurry or characterized by complex branching structures，thereby validating the effectiveness of our proposed framework for retinal vessel segmentation tasks.

With the widespread adoption of DiT (Diffusion Transformer) in controllable visual generation，the computational redundancy introduced by external control branches，such as EasyControl，and the memory⁃access bottlenecks caused by fragmented operator invocations have constrained efficient inference in large⁃scale industrial⁃grade scenarios. To address the challenges posed by multi⁃branch architectures，namely，GPU memory bandwidth limitations and graph⁃scheduling latency，we propose a software⁃hardware co‑designed，full⁃stack high⁃performance inference architecture. First，at the algorithmic level，we introduce a multi⁃branch feature‑tiling strategy combined with adaptive spatiotemporal redundancy pruning to eliminate redundant computations within the control flow. Second，at the operator level，we deeply customize global fusion operators，transforming frequent，fragmented global memory reads and writes into on⁃chip computation，thereby reducing memory⁃access overhead. Finally，at the system level，we design an FP8⁃based first⁃round calibration and static scaling quantization deployment scheme tailored for graph compilation. Experimental results demonstrate that this full‑stack acceleration approach maintains strong visual consistency across resolutions，boosting end⁃to⁃end inference throughput by approximately 2.5× on average，thus providing a viable solution for optimizing inference speed in controllable visual generation models.

To address the degradation in model performance caused by missing modalities，missing labels，and high data heterogeneity in multimodal federated learning，we propose AMRD⁃FedA²，a method that integrates relational distillation. AMRD (Adaptive Multimodal Relation Distillation) assesses the representational quality of different modalities，adaptively selects the most informative modalities，and leverages sample⁃level relation matrices to perform knowledge distillation across modalities，thereby mitigating the adverse effects of modality and label. FedA² (Federated Algorithm Adaptation Framework) adapts this distillation mechanism to various federated learning frameworks，enhancing the method's versatility in federated optimization settings. Experiments conducted on multimodal datasets，including sensor，text，image，video，and audio，using several mainstream federated learning algorithms as baselines demonstrate that，under conditions of severe modality and label，our framework outperforms the original baselines. Notably，on the Crisis⁃MMD dataset evaluated with the SCAFFOLD algorithm，the F1 score improvement exceeds 30% under the experimental conditions. Further analysis reveals that our approach performs particularly well in strongly non⁃IID (Non⁃Independent and Identically Distributed) scenarios，exhibiting robustness and strong generalization capabilities in complex，highly heterogeneous multimodal federated environments.

Event Causality Identification (ECI)，a fundamental task in natural language understanding，aims to determine causal relationships between event pairs from text and plays a crucial role in applications such as question⁃answering systems and narrative generation. Existing methods exhibit notable limitations: first，regarding the scarcity and imbalance of annotated data，incorporating external knowledge bases is prone to issues of limited coverage and noise interference; second，with respect to the blurred semantic boundaries between causal and non⁃causal samples，traditional approaches struggle to learn discriminative representations. To address these challenges，this paper proposes an ECI method based on counterfactual data augmentation and supervised contrastive learning. We design a three⁃stage data quality control mechanism “knowledge base filtering，causal structure constraint，and LLM （Large Language Model） generation and verification”. Specifically，we first extract semantically relevant intermediate concepts from structured knowledge bases. Then，we inject various foundational causal graph structures to constrain the logical coherence of generated samples. Finally，we leverage LLMs to generate and multi⁃dimensionally verify these samples，enabling the batch construction of high⁃quality，diverse counterfactual examples. Furthermore，we integrate supervised contrastive learning to explicitly optimize the semantic representation space，thereby enhancing the model’s capacity to discriminate causal boundaries. Experimental results demonstrate that our method achieves F1 scores of 74.2% and 80.4% on the EventStoryLine and Causal⁃TimeBank benchmark datasets，respectively，significantly outperforming current state⁃of⁃the⁃art models.

Image classification models face serious risks of unauthorized replication and illegal distribution. Although model backdoor watermarking techniques have been widely adopted for ownership verification，most existing methods rely on conspicuous noise triggers or rigid partitions of the interpretability space，resulting in poor watermark concealment，easy vulnerability to geometric transformations，and failure to evade manual inspection and conventional image processing. To address these issues，we propose a clean⁃label backdoor watermarking method for image classification models that leverages feature attribution and background injection. This approach achieves high concealment and strong robustness in watermark embedding without modifying the true labels of the samples. First，we employ a saliency detection mechanism to semantically decouple the foreground and background of an image，disguising the backdoor trigger as a background endowed with artistic texture features，thereby achieving visually imperceptible embedding. Next，we introduce a texture⁃adaptive superpixel segmentation strategy to establish coding targets and derive a differentiable feature attribution algorithm based on local sampling and fitting. By incorporating multi⁃task constraints，we deeply embed the watermark information into the model’s decision⁃making mechanism. Experimental results demonstrate that our proposed method not only maintains stable classification performance but also enables reliable watermark extraction and exhibits excellent resistance against common image transformation⁃based watermark removal attacks. Compared with existing watermarking schemes，our method demonstrates superior performance in terms of concealment and robustness，offering a practical copyright protection solution for ownership verification of image classification models.

Knowledge structure theory is an interdisciplinary framework at the intersection of mathematics and psychology，which has been successfully applied to personalized knowledge assessment. In multi⁃category knowledge assessment scenarios，existing research assumes that items share identical response value sets and constructs a knowledge⁃assessment automaton. However，this approach cannot accommodate more complex educational applications. Accordingly，to address multi‑category knowledge assessment settings characterized by diverse item response value sets，we develop a novel knowledge⁃assessment automaton and explore the corresponding assessment process. Specifically，we first construct a multi⁃category knowledge‑assessment automaton tailored to linear item response value sets and，based on adjacency relations among items，establish item⁃questioning rules. Next，by integrating the automaton's state⁃update mechanism with these questioning rules，we formulate a knowledge‑assessment procedure for linear item response value sets. Finally，empirical analysis demonstrates the feasibility and effectiveness of the automaton⁃based assessment process.

Cognitive state models must possess data⁃forgetting capabilities to comply with relevant laws and privacy protection requirements. However，existing cognitive diagnostic models focus on performance optimization while neglecting learners' privacy needs. To address this gap，we propose a machine⁃forgetting framework for cognitive diagnosis that tackles the problem of data forgetting in this domain. First，leveraging a partitioning approach based on cognitive states，we segment the data according to the similarity of learners' cognitive states and train cognitive diagnostic submodels in parallel within each partition. Second，employing a cross⁃attention⁃based model aggregation mechanism，we compute attention weights by fusing prediction outputs with the feature representations of the submodels，thereby enhancing the final model's generalization across diverse learner groups. Extensive experiments conducted on four real⁃world datasets demonstrate that our framework effectively enables data forgetting in cognitive diagnostic models，significantly reducing computational resource consumption and training time while maintaining model performance.

The selective hydrogenation removal of trace acetylene from ethylene⁃rich streams is critical for industrial ethylene polymerization. However，conventional supported Pd catalysts with large Pd ensembles containing contiguous active sites often suffer from low ethylene selectivity and severe over⁃hydrogenation. In this study，a series of Au@Pd/TiO₂ catalysts were prepared via a two⁃step photocatalytic deposition method. Due to the higher electronegativity of Au relative to Pd，Pd was preferentially deposited onto the surface of Au particles. Combined with characterization results，the successful construction of bimetallic catalysts with a Au core and Pd shell was confirmed. By adjusting the Pd loading，a sub⁃monolayer dispersion of Pd was achieved. In the selective hydrogenation of acetylene，the bimetallic Au@Pd catalysts exhibited significantly superior acetylene conversion and ethylene selectivity compared to the monometallic Pd catalyst.Reducing the Au particle size allowed the Pd dispersed on the Au surface to exist more as single atoms or small ensembles，thereby enhancing selectivity. In contrast，increasing the Pd loading promoted the formation of larger Pd ensembles，resulting in a sharp decline in selectivity. The optimum Au@Pd/TiO₂ catalyst achieved an ethylene selectivity of 57.8% at approximately 94% acetylene conversion，far exceeding the 88% acetylene conversion and 14.4% ethylene selectivity obtained with the monometallic Pd catalyst. This study provides a new strategy for constructing bimetallic catalysts with sub⁃monolayer dispersed Pd by tuning Au particle size and Pd loading for efficient selective hydrogenation of acetylene.