Applied Intelligence最新文献

{"title":"ValuePEFT: an effective method for steerable multi-level human values in LLMs","authors":"Jing Wang,&nbsp;Yinglin Wang","doi":"10.1007/s10489-026-07486-6","DOIUrl":"10.1007/s10489-026-07486-6","url":null,"abstract":"<div><p>The growing prevalence of Large Language Models (LLMs) necessitates methods for steering generated text toward explicitly specified human-value categories. Existing alignment and fine-tuning methods often lack a parameter-efficient interface for conditioning generation on fine-grained value labels. This paper proposes <b>ValuePEFT</b>, which maps trainable value embeddings into matrices inserted inside routed LoRA updates. We reconstruct Webis-ArgValues-22, grounded in Schwartz’s Theory of Basic Human Values, into instruction-following pairs for value-conditioned stance and premise generation. Across three random seeds, ValuePEFT achieves average stance accuracy, hard premise accuracy, and consistency accuracy of 0.634, 0.797, and 0.559, improving over the strongest controlled baselines by 3.9, 5.4, and 8.3 percentage points, respectively. Human evaluation confirms higher value correctness and stance-premise consistency without reducing fluency. Tests with paraphrased prompts, open-ended formats, and unseen conclusions show partial generalization beyond the training template, although open-ended format compliance remains challenging. These results demonstrate controllable value conditioning within the ArgValues benchmark rather than universal cultural or conflict-resolution alignment.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132923","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Self-paced deep clustering: an attention model for unsupervised domain adaptive person re-identification","authors":"Qingyu Wang,&nbsp;Ruchun Jia,&nbsp;Bo Peng,&nbsp;Xiaoyan Zhu","doi":"10.1007/s10489-026-07374-z","DOIUrl":"10.1007/s10489-026-07374-z","url":null,"abstract":"<div><p>Person re-identification (re-ID) aims at matching identity in person bounding box images extracted from non-overlapping camera views. Benefiting from the ability of supervised learning and transfer learning to extract more discriminative features, person re-ID technology has also been developed rapidly. However, methods based on supervised learning are harsh for data requirements and require precise manual annotation, which is a difficult challenge. Unsupervised learning-based methods are robust to images in complex scenes (such as low resolution and illumination angle) and do not require labeled data, so they have attracted much attention. To make up for the fact that the unsupervised learning method cannot learn the discriminative characteristics of the person, this paper proposes a new domain adaptive person re-ID model DELTA (Deep Cluster with Attention Method) based on the unsupervised learning. DELTA is more flexible and does not require additional annotations on the target domain. Furthermore, the DELTA adds a new divergence attention mechanism module (Grouped Divergence Attention Module, GDAM) that combines grouped convolution, which can capture fine-level feature information and semantic information while reducing the amount of model parameters. In addition, the DELTA adds a self-paced deep clustering module, using agglomerative clustering to cluster the fine-level and discriminative features extracted by GDAM. In this paper, rich experiments are carried out on the Market1501, DukeMTMC-reID and MSMT17. DELTA obtains best result on rank-1, rank-5 and mAP, thus verifying the effectiveness of the proposed DELTA.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132644","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GDPB: a gradient-decoupled privileged baseline for UAV navigation under compound disturbances","authors":"Xin Wang,&nbsp;Jiankang Zhao","doi":"10.1007/s10489-026-07533-2","DOIUrl":"10.1007/s10489-026-07533-2","url":null,"abstract":"<div><p>UAV navigation in real-world environments is often affected simultaneously by wind fields, actuation noise, control delay, and sensor degradation. Under compound disturbances, training samples differ substantially in their states and environmental conditions, making a single batch-mean return an inadequate reference for the expected return at each state. We propose the Gradient-Decoupled Privileged Baseline (GDPB), which predicts Monte Carlo returns from onboard observation features and simulator-provided privileged disturbance parameters to supply a state-dependent baseline for policy updates. Stop-gradient is applied to the baseline prediction when computing advantages, preventing the policy loss from directly updating the baseline branch; the branch is trained separately through return regression. At deployment, the baseline branch and privileged inputs are removed, and the policy relies only on onboard observations. On a compound-disturbance navigation task in Isaac Sim, GDPB achieves a composite success rate of 96.47%, outperforming PPO, GRPO, SAC, and TD3 with a smaller standard deviation across repeated runs. Its advantage over PPO is more pronounced under moderate-to-strong disturbances. These results show that GDPB improves navigation robustness and training stability without adding deployment-time computation.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132620","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DPAS-Net: dynamic prototype alignment and separation network for incomplete data multi-view clustering","authors":"Yufeng Sun,&nbsp;Changming Zhu","doi":"10.1007/s10489-026-07510-9","DOIUrl":"10.1007/s10489-026-07510-9","url":null,"abstract":"<div><p>Incomplete multi-view clustering (IMVC) aims to identify a consistent cluster structure when some views are unavailable. In prototype-based deep IMVC, prototype estimates derived from individual mini-batches may fluctuate with the distribution of observed samples and views, while insufficient regularization of inter-prototype geometry can reduce cluster discrimination and weaken the reliability of prototype-guided completion under severe incompleteness. To address these limitations, we propose the Dynamic Prototype Alignment and Separation Network (DPAS-Net), which coordinates representation learning, prototype regularization, and latent missing-view completion. Cross-view attention and fusion-guided view-specific reconstruction are first employed to integrate complementary information while preserving view-dependent characteristics. A robust instance-level consistency objective then aligns jointly observed samples. At the cluster level, soft sample-to-prototype assignments generate mini-batch prototype estimates, which are progressively aggregated into a shared prototype matrix. An explicit separation regularizer further discourages excessive similarity among distinct prototypes. During the final fine-tuning epoch, view-specific prototype banks are constructed from observed embeddings and retained for post-training latent missing-view completion. Experiments on four benchmark datasets, five MCAR missing rates, and five training seeds show that DPAS-Net achieves the highest mean ACC, NMI, and F-score in all 60 reported dataset–missing-rate–metric settings among 12 representative baselines. At a missing rate of 0.9, DPAS-Net obtains ACC values of <span>(88.50pm 0.65%)</span>, <span>(94.95pm 0.28%)</span>, <span>(42.89pm 0.43%)</span>, and <span>(74.17pm 0.73%)</span> on Caltech101-7, HandWritten, Scene-15, and ALOI-100, respectively, exceeding the strongest competing methods by 3.11, 6.53, 2.27, and 4.15 percentage points. Seed-matched tests provide exploratory evidence that many of these gains are stable across the evaluated seeds, while the ablation results support the complementary contributions of the coupled components. Additional evaluations under MAR and MNAR show that performance generally decreases under structured missingness, with severe MNAR posing the greatest challenge.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132493","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"RPCV: robust prototype alignment via cross-view attention for incomplete multi-view clustering","authors":"Hengbin Li,&nbsp;Changming Zhu","doi":"10.1007/s10489-026-07508-3","DOIUrl":"10.1007/s10489-026-07508-3","url":null,"abstract":"<div><p>Incomplete Multi-View Clustering (IMVC) aims to uncover consistent clustering structures from data with missing views. Recently, prototype-based deep clustering methods have garnered significant attention owing to their capacity to capture the global semantic structure of data. However, existing prototype-based multi-view methods typically rely on the assumption that features extracted by front-end encoders are of high quality, which renders prototype generation susceptible to noise. Furthermore, when addressing the prototype-unaligned problem (PUP), these methods often place excessive emphasis on inter-view consistency, thereby overlooking the complementary and diverse information present across different views. To address these challenges, this paper proposes a Robust Prototype Alignment via Cross-View Attention for Incomplete Multi-View Clustering (RPCV). Specifically, we design a transformer-based cross-view attention module within the feature extraction stage. This module leverages the contextual information from observed views to dynamically guide inference of latent representations of missing views, effectively suppressing feature noise and enhancing the discriminability of latent representations. Then, we use confidence-weighted interpolated relational consistency learning to mitigate the heterogeneity of multi-view data. Subsequently, we introduce a robust prototype learning via dual optimization module, which innovatively incorporates a joint entropy constraint during the prototype relation learning phase. This mechanism strikes a balance between maintaining cross-view semantic alignment and preserving inter-view differences. Finally, we employ a prototype-based imputation strategy to recover the missing multi-view data. Extensive experimental results demonstrate that RPCV achieves superior performance on clustering tasks compared to various existing incomplete multi-view methods. The code is available at: https://github.com/YDSHlee/rpcv.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132449","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DSGAFormer: decoupled static-dynamic graph-augmented transformer for traffic flow forecasting","authors":"Yuxi Feng,&nbsp;Lian Xiong,&nbsp;Deliang Li","doi":"10.1007/s10489-026-07513-6","DOIUrl":"10.1007/s10489-026-07513-6","url":null,"abstract":"<div><p>Traffic flow forecasting requires modeling both time-varying traffic patterns and fixed road-network topology. Existing methods commonly fuse static topology and dynamic temporal features in a shared representation space, which may introduce optimization interference and weaken the preservation of topological priors. To address this issue, we propose DSGAFormer, a Decoupled Static-Dynamic Graph-Augmented Transformer with a Subspace Overwriting Mechanism. Parallel temporal and spatial self-attention branches first capture global dependencies and are integrated through cross-attention. The static branch uses one-time Bernstein spectral initialization to construct topology-informed node embeddings, whereas the dynamic branch extracts temporal contexts using a lightweight two-layer multilayer perceptron. Mixed Graph Cross-Attention subsequently aligns the static and dynamic contexts with the backbone representation. The resulting graph-augmented features overwrite only the reserved adaptive channels, leaving the core channels unchanged at the fusion point. Experiments on four public PEMS datasets demonstrate competitive forecasting performance. In particular, DSGAFormer achieves an MAE of 14.51 and an RMSE of 24.32 on PEMS03, reducing RMSE by 11.5% relative to ST-MambaSync. Ablation studies further demonstrate the contributions of static topology, dynamic context, and subspace overwriting.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132109","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"ARCHEX: explaining models through archetypal decomposition for discovering complementary patterns in model interpretability","authors":"Abraham Itzhak Weinberg","doi":"10.1007/s10489-026-07506-5","DOIUrl":"10.1007/s10489-026-07506-5","url":null,"abstract":"<div><p>While local explainability methods like SHAP and LIME have become standard tools for model interpretation, they often miss the global structural patterns that define data archetypes. We introduce <b>ARCHEX</b> (ARCHetype-based EXplainer), a simplex-constrained archetypal analysis framework that identifies interpretable extreme data profiles and uses them, via a simple linear surrogate trained on the original prediction target, to approximate a trained model’s decision boundary from a compressed, <i>k</i>-dimensional representation. We characterize precisely where ARCHEX’s sparsity comes from: because archetype membership weights lie on the probability simplex, their <span>(ell _1)</span> norm is constant, so sparsity arises from simplex projection itself rather than from <span>(ell _1)</span> regularization. We evaluate ARCHEX on five public benchmark datasets spanning four domains, using adaptive <i>k</i>-selection restricted to the training partition, a linear (non-black-box) surrogate, and bootstrap confidence intervals on both predictive performance and the rank correlation between ARCHEX’s archetype-derived feature ranking and SHAP’s local attribution ranking. This rank correlation is small in magnitude and, once its sampling uncertainty is quantified, statistically indistinguishable from zero in four of five datasets; in the one dataset where the interval excludes zero (Digits), the correlation is moderate and positive rather than low, arguing against treating a low point estimate as established evidence of complementary explanatory content. Independent of this, ARCHEX’s soft, convex archetype membership outperforms a hard <i>k</i>-means-cluster-membership ablation on held-out predictive accuracy in all five datasets. On the Breast Cancer Wisconsin dataset, ARCHEX’s top-ranked features are concavity- and concave-points-related measurements, several of which SHAP ranks far lower; we present this as a preliminary, descriptive observation rather than a validated clinical finding. We position ARCHEX as a global data-structure and subgroup-discovery method that complements, rather than replaces, local attribution methods.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132492","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"CLARiF: clue alignment and retrieval-informed fusion for fine-grained pest recognition","authors":"Qiuyu Li,&nbsp;Yushan Pan,&nbsp;Nan Xiang,&nbsp;Haiyang Zhang,&nbsp;Zuhe Li,&nbsp;Xiaowei Huang,&nbsp;Zhijie Xu","doi":"10.1007/s10489-026-07522-5","DOIUrl":"10.1007/s10489-026-07522-5","url":null,"abstract":"<div><p>Fine-grained pest recognition in field imagery remains difficult because of cluttered backgrounds, life-stage variation, subtle inter-class differences, long-tailed categories, and unsupported taxa. We propose CLARiF, a vision–language framework that combines structured diagnostic captioning, mask-guided part extraction, bidirectional part–attribute fusion, and exemplar-informed re-ranking. On the official IP102 test split, CLARiF achieves 77.81 ± 0.18% accuracy and 77.07 ± 0.21% macro-F1. Under comparable data-processing and fine-tuning settings, it also outperforms independently reproduced vision–language baselines while retaining each model’s native language backbone. A fixed held-out-species protocol yields 88.62% AUROC for unsupported-input rejection.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"MAF-Net: multi-scale attention fusion network for pancreas segmentation","authors":"Yang Li,&nbsp;Hongwei Deng","doi":"10.1007/s10489-026-07514-5","DOIUrl":"10.1007/s10489-026-07514-5","url":null,"abstract":"<div><p>Pancreas segmentation is a critical task in abdominal CT image analysis; however, achieving precise segmentation remains challenging due to the pancreas’s variable shape, ambiguous boundaries, and low contrast with surrounding tissues. To address these issues, we propose a novel pancreas CT image segmentation network called MAF-Net. This network introduces the Multi-scale Fusion with Spatial Attention block (MFSA block), which employs parallel multi-scale convolutions to extract contextual information at different scales while incorporating a spatial attention mechanism to enhance structural awareness and localization capabilities for the pancreatic region. Furthermore, we design an Attention Fusion Module (AFM) that dynamically allocates appropriate weights between local details and global semantics, significantly improving the overall feature representation of the pancreas. Experiments conducted on the NIH public dataset demonstrate that our method achieves an average Dice Similarity Coefficient (DSC) of 85.47±4.64%, representing a 2.19% improvement over baseline models and outperforming other segmentation models.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132076","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Multi-granularity contrastive learning for discovering categories of teacher instructional behavior indicators","authors":"Ting Cai,&nbsp;Qingyuan Tang,&nbsp;Yu Xiong,&nbsp;Lu Zhang,&nbsp;Attila Pandur","doi":"10.1007/s10489-026-07420-w","DOIUrl":"10.1007/s10489-026-07420-w","url":null,"abstract":"<div><p>In complex and dynamic educational environments, the discovery of categories of teacher instructional behavior indicators is crucial for teaching evaluation. Generalized Category Discovery (GCD) offers a novel approach for automatically uncovering such indicators from teaching videos. However, existing methods suffer from inadequate capture of spatiotemporal dynamics and semantic interference in complex classroom environments, which limits the effectiveness of novel category discovery. This paper proposes the MGCL model, a multi-granularity contrastive learning framework. Guided by known indicator category data, it mines large-scale unlabeled data to achieve automated identification and discovery of new categories for instructional behavior indicators. In this framework, instance-level contrastive learning (IL-CL) optimizes the characteristic representation of individual samples; neighborhood consistency contrastive learning (NC-CL) enhances the discriminability of samples within local contexts by emphasizing the semantic consistency of neighboring samples; and global category consolidation contrastive learning (GC-CL) leverages pseudo-labels to optimize category partitioning at a holistic level, thereby improving inter-class separation. Operating at the instance, neighborhood, and global levels respectively, these components work in concert to collectively enhance feature discriminability and category stability. Experiments were conducted on the proprietary TBU dataset and the public UCF101 and HMDB51 datasets. Results demonstrate that the proposed method achieves state-of-the-art performance across multiple metrics. This study provides a feasible and effective solution for discovering categories of instructional behavior indicators in complex educational environments.</p></div>","PeriodicalId":8041,"journal":{"name":"Applied Intelligence","volume":"56 15","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149132075","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}