Large language models (LLMs) can translate natural language into optimization code, but silent failures pose a critical risk: code that executes and returns solver-feasible solutions may encode semantically incorrect formulations – a feasibility-correctness gap reaching 90 percentage points on compositional problems. We introduce ReLoop, which addresses this gap through two complementary mechanisms. Structured generation decomposes code production into a four-stage reasoning chain (understand, formalize, synthesize, verify), preventing formulation errors at their source. Behavioral verification detects errors that survive generation by testing whether the formulation responds correctly to solver-based parameter perturbation – an external semantic signal that bypasses LLM self-review and requires no ground truth. The two mechanisms are complementary by error structure: structured generation drives the largest gains on compositional problems (+8.5pp accuracy on RetailOpt-190 with Claude Opus 4.6), while behavioral verification dominates on localized defects (+4.4pp on MAMO-ComplexLP, its largest contribution across benchmarks). Combined with diagnostic execution recovery, ReLoop reaches 100

Interstitial lung diseases (ILDs) are a heterogeneous group of pulmonary disorders where delayed risk stratification significantly increases mortality. However, accurate early prognosis is often hindered by highdimensional, redundant clinical features and the complex non-linear interactions among them. We propose a hybrid feature selection model for ILD mortality prediction. WBCPSMA, a novel optimizer integrating Comprehensive Learning Particle Swarm Optimization (CLPSO) and the Slime Mould Algorithm (SMA) with adaptive weighting exploration and optimal backtracking guidance, was combined with a Kernel Extreme Learning Machine (KELM) classifier to enable accurate prediction. The method was validated on the CEC2014 and CEC2017 benchmark suites to confirm its global optimization capability and evaluated on a clinical ILD dataset comprising 211 patients and 22 features. To ensure rigorous validation on a limited dataset, we employed a Nested Cross-Validation protocol with Bootstrap resampling. The model achieved an Accuracy of 0.806 (95% CI: 0.742-0.855) and an AUC of 0.836, outperforming seven state-of-the-art methods. Crucially, it demonstrated superior clinical utility by maintaining a balanced Sensitivity (0.798) and Specificity (0.814) via dynamic thresholding, alongside reliable probability calibration (Brier Score: 0.162). The algorithm identified four key biomarkers-HRCT score, PaO2/FiO2, Lactate, and Age-with SHAP analysis confirming their clinical relevance. Furthermore, external validation on four UCI datasets verified the method's robust generalization capability. This approach offers high predictive performance and transparent decision support for ILD mortality risk stratification, providing a foundation for future integration into clinical workflows.

The RIME algorithm, a physics-inspired metaheuristic algorithm proposed in recent years, is recognized for its simplicity, minimal parameter requirements, and rapid convergence. However, it faces challenges such as premature convergence to local optima and limited solution precision in practical applications. To address these limitations, this study introduces HDRIME, a multi-strategy enhanced variant of RIME. HDRIME was rigorously evaluated using the IEEE CEC 2017 benchmark suite. The preliminary analysis focuses on the impact of its introduction strategy on the overall algorithm and historical search trajectories, demonstrating its enhanced capability to balance global exploration and local exploitation. The stability test further revealed the robustness of HDRIME to the change of problem dimensions. Extensive comparative evaluations with 7 baseline algorithms and 8 cutting-edge variants validate the superiority of HDRIME across diverse optimization scenarios, especially confirming its advantages and potential for high-dimensional problems. To validate its practical utility in high-dimensional scenarios, a binary version, bHDRIME, was developed for feature selection tasks. Evaluations on 72 public datasets, including 66 high-dimensional microarray gene expression datasets—demonstrated that bHDRIME achieves higher classification accuracy with significantly fewer selected features compared to 7 established algorithms. Notably, bHDRIME also substantially reduced computational time compared to its predecessor, bRIME. In conclusion, HDRIME exhibits exceptional performance across problems of varying dimensions, with competitiveness in high-dimensional optimization, especially for feature selection in microarray gene expression datasets. The code and data are available as open source at: https://github.com/AlexandreShogenji/HDRIME.

Variations in tissue morphology and image quality make threshold selection challenging. Furthermore, traditional pathological diagnosis relies heavily on expert interpretation and is subject to subjective variability, which limits diagnostic accuracy and efficiency. Therefore, this study proposes an improved AO (DCAO) incorporating the random follower search (RFS) strategy and the crisscross optimization strategy (CC). The RFS enhances inter-population communication by reducing the algorithm’s reliance on the current global best individual. The CC improves population diversity, facilitates escaping local optima, and strengthens the algorithm’s global search capability. DCAO’s performance was validated in two parts. First, it was benchmarked against 24 algorithms on the IEEE CEC 2014 and 2022, where statistical analysis revealed its robust competitiveness. Second, when applied to breast cancer histopathological segmentation against eight algorithms, across all tested threshold levels, DCAO ranked first on FSIM, PSNR, and SSIM—improving them by 1.03%, 2.70%, and 0.66% over the second-best method, respectively. It also achieved the highest 2D Kapur entropy, the fastest convergence, and demonstrably superior visual quality with clearer boundaries. This method provides a reliable and efficient solution for cancer image analysis, offering technical support for accurate lesion localization and treatment planning. However, it has limitations: the introduction of strategies increases computational complexity; the algorithm is designed only for 2D images; and its robustness against staining variations has not been systematically evaluated.

The critical role of accurate segmentation in breast cancer diagnosis motivates this research. We propose a novel approach that integrates an advanced optimization algorithm to significantly improve the segmentation precision of pathological images. Specifically, we first improve the Differential Evolution (DE) algorithm by incorporating the Population Information Exchange (PIE) strategy and the Covariance Matrix Adaptation Evolution Strategy (CMAES) to propose TDE, The PIE mechanism improves the global search ability of the algorithm through random population information exchange, effectively suppressing premature convergence. The CMAES strategy generates high-quality solutions based on covariance learning, significantly improving the accuracy of local exploitation. Subsequently, TDE is coupled with Renyi entropy, forming the core of our segmentation methodology for breast histopathology images. We initially verified the optimization capability of TDE on the IEEE CEC 2017 functions set. Statistical validations through the Wilcoxon signed-rank test and Friedman test confirmed the superiority of TDE. Following this, we evaluated the multi-threshold segmentation model combining TDE and Renyi entropy on 9 breast cancer pathology images, demonstrating that the TDE algorithm outperforms peer algorithms. The comprehensive experimental results affirm TDE's outstanding performance in both optimization benchmarks and medical image segmentation applications.

In the realm of online education, it is a fundamental task to assess the mastery of knowledge points by students and to customize personalized exercises for them. However, recommending suitable exercises for students poses a challenge due to the varying levels of student knowledge and the extensive exercise question bank. Current methodologies utilize collaborative filtering to recommend exercises, which lack consideration of changes in student knowledge, making it difficult to capture student behavior and utilize the relationship information between students and exercises. Therefore, this paper proposes an end-to-end knowledge graph (KG) enhanced exercise recommendation algorithm for multi-task learning (KERM). By incorporating a feature sharing unit, the recommendation task and knowledge graph embedding (KGE) task automatically share additional feature information with each other, thereby enhancing the accuracy of exercise recommendation. Firstly, a self-adaptive adjustment factor has been designed that continuously monitors the changes in knowledge state of the students in the recommendation task. Subsequently, a cleverly designed single multi-level feature interaction is implemented in the KGE task where single-level feature interactions extract detailed entity information and capture rich interaction details within entity neighborhoods, multi-level feature interactions aggregate multiple single-level interactions to expand receptive field coverage and improve quality of feature interaction. Finally, a feature sharing unit is designed to model high-order interactions between student and exercise features by automatically sharing additional information between both tasks to help prevent overfitting while improving robustness. To verify the effectiveness of KERM, extensive experiments are conducted on four real datasets yielding average prediction accuracy by 4.2

Hyperspectral images exhibit strong spectral continuity, resulting in significant inter-band correlations that make many bands redundant for practical applications. While band selection methods can improve efficiency through dimensionality reduction, conventional approaches relying on distance-based spectral relationship evaluation suffer from exponential computational complexity and tend to produce suboptimal solutions due to their limited ability to distinguish similarly distanced bands. This paper presents HSI-SSRUN, a novel framework combining a multi-strategy spectral relationship representation mechanism with an enhanced evolutionary algorithm. The proposed method first constructs a 3D spectral relationship space by integrating Kapur’s entropy and JM distance, effectively capturing pixel-level information content and contextual band correlations. Second, it introduces an evolutionary optimizer that balances global exploration and local exploitation through hybrid foraging strategies, ensuring robust convergence while avoiding local optima in nonlinear spectral relationship spaces. Extensive experiments demonstrate the framework's reliability, efficiency, and stability, confirming its effectiveness in supporting downstream classification tasks.

Path planning is vital for Unmanned Aerial Vehicle (UAV), especially in low-altitude environments. This paper proposes Interactive Mechanism and Enhanced Solution Quality-driven Artemisinin Optimization (IEAO) to address UAV path planning challenges. The Interactive Mechanism enhances population diversity through communication among individuals within the population. Furthermore, IEAO combines the Artemisinin Optimization update rule with an Enhanced Solution Quality strategy to refine solution selection, thereby enhancing convergence accuracy. The effectiveness of IEAO was evaluated using IEEE CEC2017 and IEEE CEC2019 test sets. The results show that IEAO outperforms other classical and advanced algorithms in 87.26

This paper presents an efficient and effective variant of the competitive swarm optimizer (CSO) named multi-search pattern CSO (MSPCSO). Two tailored strategies are integrated into MSPCSO to enhance the robustness of CSO: the multi-search pattern mechanism aims to enrich the search diversity and further maintain the population diversity during optimization, and the success history-based parameter tuning strategy allows MSPCSO to adapt complex fitness landscapes in various optimization challenges. We conduct optimization experiments in CEC benchmarks and engineering challenges against nine cutting-edge optimizers to investigate the performance of MSPCSO. Rigorous statistical tests confirm the superiority and competitiveness of MSPCSO, while the sensitivity and ablation experiments are implemented to comprehensively analyze the performance of MSPCSO. Furthermore, we extend MSPCSO to ensemble learning methodology for the coffee leaf disease diagnosis, where several pre-trained deep learning models are trained in the coffee leaf disease dataset. The top three deep learning models are employed for ensemble learning according to the MSPCSO-optimized soft voting scheme. Experiments in public datasets confirm that the proposed MSPCSO-Ensemble achieves significant improvements in the accuracy of 1.021

Experience-learning agents for optimization modeling improve by storing verified skills, but existing learners admit knowledge by checking against known answers, which real ticket streams do not provide. The natural label-free alternatives are unreliable: on a 300-problem label-blind stream, admitting every executable model poisons roughly one admission in four, while single-instance agreement accepts models that match at one value but differ elsewhere. We propose AdmitOR, an admission gate built on calibrated external behavioral evidence. Candidates from three model families, prompting strategies, and solver stacks are run on instances resampled from an extracted parameter domain; agreement across the resulting value-function traces is summarized by a cross-family clique, and a calibrated threshold returns accept, abstain, or escalate. The preregistered false-discovery criterion holds on calibration data but not on the wild stream. We report this negative result in full and trace most failures to benchmark texts that do not faithfully encode their labeled instances. Comparing four admission judges on one collection of logs inside a state-of-the-art skill learner, AdmitOR raises admission precision to 0.927, against 0.871 for majority vote and 0.726 for execution success, yielding 3.1x and 8.0x fewer poisoned admissions. Its library is the smallest and attains the highest macro accuracy across five public benchmarks, 58.4 against 54.8 for majority vote and 53.9 for the ground-truth-labeled library. The 3.5-point gain over majority vote is supported by a paired bootstrap and survives correction for a host-side anomaly. To our knowledge, AdmitOR is the first label-free admission mechanism designed around an explicitly calibrated false-discovery target. The transfer failure identifies a necessary condition for extending it to wild streams.

In the field of Chinese medicine, accurate discrimination of decoction pieces is crucial for ensuring the quality and efficacy of traditional Chinese medicine (TCM). However, meeting the demands of modern TCM development is challenging due to the variety and shapes of decoction pieces and the low efficiency of traditional identification methods. This study proposes the multi-scale lightweight network model, LW-MobileNetV2, as an efficient and suitable solution for resource-constrained mobile devices to identify TCM pieces. Based on MobileNetV2, a multi-scale feature extraction layer is created using group convolution, followed by the compression of the model. Additionally, the h-swish activation function is incorporated for enhancing the model's ability to recognize Chinese herbal pieces. A small dataset comprising 17 types of classic TCM decoction pieces, which are difficult to distinguish, was created by capturing images to provide data support for subsequent research on model image recognition. According to the experimental results, the LW-MobileNetV2 attains an average precision rate of 93.716 %, a recall rate of 93.900 %, and an F1-Score of 93.622 % on the test set, with an FPS of 281.23, and the parameter memory being 4.85 MB, which is 43.4 % smaller than that of MobileNetV2. In contrast to the classic residual network ResNet18 and other lightweight networks such as GhostNet, ShuffleNet, and MobileNetV3, LW-MobileNetV2 possesses evident strengths in recognition accuracy, speed, and parameter memory. In comparison with the original model, the computational cost is decrease to 1.99E + 08, and the amount of memory required is reduced from 8.57 MB to 4.85 MB. The model possesses strong practicability and is capable of meeting the demand of resource-limited mobile devices for the recognition of TCM decoction slices.

This study developed a machine learning framework for stratifying abnormal glucose tolerance risk among outpatients presenting with foamy urine. It retrospectively analyzed 64 features from 511 patients with foamy urine in the outpatient medical record system. Patients were divided into two groups based on oral glucose tolerance tests (OGTT) results. This study proposed a novel predictive framework, bPKPLO-SVM, which integrates the Prism-refraction K-means Polar Lights Optimizer (PKPLO) with the Support Vector Machine (SVM) for feature selection and early diabetes diagnosis. The proposed PKPLO algorithm enhances the Polar Lights Optimizer (PLO) by incorporating mechanisms for improved search diversity and solution refinement, inspired by Prism refraction and K-means clustering. Results showed that bPKPLO-SVM performed well in early diabetes prediction, achieving an accuracy of 81.639% and a recall of 85.927%, outperforming many common prediction models and similar algorithms. Model analysis further identified fasting plasma glucose (FPG), glycosylated hemoglobin (HbA1c), platelet distribution width-standard deviation (PDW-SD), urine specific gravity (USG), and age as frequently selected predictors. These findings suggest that bPKPLO-SVM may serve as a supportive risk-stratification tool for patients presenting with foamy urine, rather than a universal screening model for the general population.

The continuous version of the Ant Colony Optimization algorithm (ACOR) is a meta-heuristic method inspired by information exchange in ant colony. While ACOR is widely recognized for its simplicity, ease of implementation, and competitive performance in low-dimensional optimization problems, it faces significant challenges in high-dimensional and complex scenarios. These challenges include low convergence precision, slow convergence speed, and a tendency to fall into local optima. This study proposes a novel robust and adaptive ant colony optimization algorithm called Elite Guidance Ant Colony Optimization (EGACO) to address these limitations. EGACO introduces an elite guidance strategy that allows elite populations to guide the updates of ordinary populations while maintaining diversity and robustness. Additionally, EGACO replaces Gaussian sampling with adaptive t-distribution sampling, dynamically adjusting the degrees of freedom to ensure a more balanced search range. To comprehensively evaluate the performance of EGACO, we conducted extensive experiments on CEC2017 benchmark functions with 30, 50, and 100 dimensions, comparing it against ACOR and 15 other state-of-the-art algorithms. Furthermore, EGACO was applied to multi-threshold segmentation of colon cancer histopathology images. The results show that EGACO has a significant advantage over ACOR on 29 of the CEC2017 functions (p ≤ 0.05) and maintains superior performance compared to other leading algorithms. It also leads in both quantitative and qualitative segmentation results, highlighting its potential for solving complex real-world problems.

Iris localization and segmentation constitute mission-critical preprocessing stages in iris recognition systems, where their precision directly governs overall recognition accuracy. However, iris images captured under noncooperative conditions are prone to boundary distortions caused by eyelash or eyelid occlusions and defocus blurring, while texture features suffer from weakened saliency due to uneven illumination or specular reflections, leading to reduced algorithm robustness. To address these challenges, this paper proposes a cascade attention feature residual fusion network (CA-RFNet) for multitask iris localization and segmentation in unconstrained scenarios. CA-RFNet adopts an encoder-decoder structure with skip connections. In the encoder stage, deep convolutional residual blocks hierarchically extract iris texture features. A cascade attention fusion module embedded in skip connections dynamically weights and adaptively integrates multi-receptive-field features while enabling cross-scale information complementarity. The decoder incorporates a boundary perception module with cross-layer feature interaction mechanisms to enhance fine-grained structural perception and cross-hierarchy semantic representation, thereby improving edge prediction accuracy. CA-RFNet modules work collaboratively to overcome adverse effects of unconstrained subject behaviors and complex environmental interference on algorithm robustness in non-cooperative scenarios. Extensive experiments on five non-cooperative iris datasets (CASIA-Iris-Distance, CASIA-Iris-Complex-Occlusion, CASIA-Iris-Complex-Off-angle, CASIA-Iris-M1, and CASIAIris-Africa) demonstrate that CA-RFNet achieves superior segmentation and localization performance on challenging samples with complex noise factors including occlusion, off-angle, illumination variation, specular reflection, dark iris, and dark skin.

Despite the improvement of survival outcomes in patients with locally advanced non-small cell lung cancer (NSCLC) due to neoadjuvant immunochemotherapy and the following radical resection, the incidence of postoperative early recurrence (ER) within this cohort remains alarmingly high, ranging from 20 % to 40 %, including those who have achieved a major pathological response, where rates span between 10 % and 20 %. Consequently, identifying postoperative ER to guide postoperative management is imperative for augmenting therapeutic outcomes and patient survival quality and duration. This study aims to develop a machine-learning model to forecast postoperative ER in NSCLC patients with neoadjuvant immunochemotherapy. Eligible stage IIB-IIIB NSCLC patients receiving neoadjuvant immunochemotherapy across four high-volume medical centers were retrospectively reviewed. Clinicopathological features and survival outcomes were collected. A machinelearning model named bDIINFO-SVM, integrating a dual-population collaborative-driven weighted mean of vectors (DIINFO) algorithm with a support vector machine (SVM) classifier, was developed. This study included 173 cases with a median follow-up duration of 43.0 months. The optimization capabilities of the DIINFO algorithm are demonstrated through comparison with a set of state-of-the-art algorithms under the CEC 2017 benchmark function. Further comparative analysis revealed that the bDIINFO-SVM model exhibited a significant advantage in accurately predicting the ER of NSCLC patients, compared to other peer models and well-known classifiers, achieving an accuracy of 98.856 % and a sensitivity of 99.091 %. This study offers an effective tool for predicting the postoperative ER of NSCLC patients with induction immunochemotherapy, potentially impacting the formulation of postoperative follow-up and adjuvant treatment strategies positively.

Qubit loss is a major source of error in quantum computation, as it invalidates the algebraic structure of the standard stabilizer formalism for quantum error-correcting codes. On the one hand, it complicates decoding; on the other hand, it introduces stochastic flicker patterns in stabilizers as a hallmark of qubit loss. Here, we develop an artificial-intelligence-enabled decoder based on a spatiotemporal Graph Neural Network (STGNN) architecture to extract spatial and temporal correlations from syndrome histories. Our decoder performs a dual-head task, simultaneously correcting standard Pauli errors and identifying the locations of qubit loss. Our decoder achieves significantly higher logical accuracy than both the traditional minimum-weight perfect matching (MWPM) algorithm and even delayed-erasure MWPM decoders that use qubit loss information from the final round as input. Our decoder can also identify more than 90

Feature selection can improve predictive accuracy and reduce computational cost in high-dimensional data analysis. However, wrapper-based methods may converge prematurely when the associated discrete search space is large. This study integrates a near-distance to multi-target (NDM) strategy and a rebound boundary handling (RBH) mechanism into the polar lights optimizer (PLO), yielding NPLO. NDM introduces a proximity-aware auxiliary objective for environmental selection, whereas RBH preserves directional information through mirror reflection at a violated bound. On the 12 IEEE CEC 2022 functions, NPLO obtained the best Friedman average ranks of 2.5833 against 11 baseline algorithms and 1.7500 against 11 enhanced competitors, although alternative algorithms remained superior on individual functions. Search-trajectory and population-diversity analyses further characterize its exploration–exploitation behavior. A binary implementation, bNPLO, was evaluated for feature selection on 24 datasets with eight established binary metaheuristics. It achieved the lowest or tied-lowest classification error on most datasets while generally retaining compact feature subsets. These results support the competitiveness of NPLO on the evaluated continuous and discrete problems.

In real clinical settings, medical image datasets are often partially annotated due to high labeling costs and complexity, which limits multi-label classification. Existing methods often attempt to tackle this challenge by either decoupling features to generate pseudo-labels or by treating all unknown labels as negative labels during training. In scenarios with severe label scarcity, the former approach may fail to generate high-quality decoupled features, while the latter is prone to introducing label noise. To address these challenges, we propose a novel method for partial multi-label medical image recognition tasks, called Asymmetric Dual Thresholds and Co-occurrence Relationship (ADTCR). Specifically, ADTCR consists of two pseudo-label generation strategies: Asymmetric Dual Threshold (ADT) and Co-occurrence Relationship (CR). The ADT strategy is designed to initially identify pseudo-labels by applying a lower threshold for negative pseudo labels and a higher threshold for positive pseudo labels, ensuring the generation of high-quality pseudo labels. Meanwhile, the CR strategy aims to uncover potential positive labels by capturing label co-occurrence relationships, enabling the detection of latent positive labels among the unknown ones. Finally, to assess the model's confidence in the generated pseudo-labels, we design a Threshold-based Weighting Loss (TWL), which uses threshold-based weights to weight each pseudo-label, thereby further improving performance. Extensive experiments conducted on three multi-label medical image datasets, i.e., Axial Spondyloarthritis, NIH Chest X-ray 14, ODIR-5K, demonstrate that our method achieves state-of-the-art performance.

The digital transformation of medical education represents a significant global development trend. However, clinical skill assessment, a vital method to inspect education outcome, has historically depended on conventional manual evaluation methods. To further explore intelligent training paradigms and realize autonomous instruction and automated assessment, this study introduces contrastive learning methods to construct a novel clinical skill assessment system characterized by unified standard, visible process, and objective evaluation. Practical validation indicates that the model effectively differentiates varying levels of proficiency in the specific surgical procedures. Under the optimal parameter combination, the prediction of one operational procedure achieved a peak accuracy of 94.01%. With a one-point tolerance, accuracy further improved to 96.41%. Overall, depending on its objectivity and standardization, the model significantly enhances the scientific rigor and efficiency of the clinical skill training process, while also providing immediate feedback on performance.