Intermittent demand forecasting faces challenges including sparse observations, cold starts for new items, and product discontinuation. Classical methods (e.g., Croston, TSB) lack generative foundations, while deep learning models require large-scale data and suffer from poor interpretability. This paper proposes TSB-HB—a hierarchical Bayesian model that jointly models demand occurrence (via Beta-Binomial) and nonzero demand size (via Log-Normal), leveraging hierarchical priors to enable partial pooling across items. This design preserves item-level heterogeneity while enhancing robustness in estimating sparse demand sequences. TSB-HB delivers calibrated probabilistic forecasts and maintains strong interpretability. Evaluated on the UCI Online Retail dataset and an M5 subset, it significantly outperforms baselines—including Croston, SBA, and TSB—achieving substantial reductions in RMSE and RMSSE. These results demonstrate its generalizability and practical utility for intermittent demand forecasting.

This work addresses the limited diversity and mode collapse prevalent in adversarial prompt generation for red-teaming large language models (LLMs). We propose the first fine-grained probabilistic generation framework for attack prompts based on Generative Flow Networks (GFlowNets). Our method mitigates mode collapse via smoothed training and jointly optimizes attack diversity and success rate by integrating toxicity classifier guidance and cross-model transfer training. The generated prompts exhibit strong generalization and transferability across diverse target LLMs—including both safety-finetuned and base models. When used to construct a red-teaming dataset, our prompts significantly improve the robustness of safety-finetuned models against other reinforcement-learning-based red-teaming attacks (+37% success reduction). This work establishes a scalable, high-coverage paradigm for automated, probabilistic adversarial prompt generation, advancing systematic and rigorous safety evaluation of LLMs.

Visual generative models often exhibit unstable and error-prone responses to complex concepts, yet existing work lacks systematic analysis and effective mitigation strategies. To address this, we propose Online Concept Balancing (OCB), a training-time intervention that introduces a concept-level imbalance-aware loss (IMBA Loss) to dynamically calibrate response biases—without requiring offline data resampling or architectural modifications. Through causal-driven empirical analysis, we identify root causes of concept imbalance and construct Inert-CompBench, a novel benchmark, alongside a multi-source test suite for rigorous evaluation. On three public benchmarks, OCB significantly improves baseline models’ accuracy on complex concepts and compositional robustness, yielding an average gain of +12.7%. The method demonstrates strong generalization across architectures and datasets, and operates as a plug-and-play module with minimal implementation overhead.

Detecting anomalous behaviors in densely crowded scenes remains challenging due to severe occlusion and context-dependent motion patterns. To address this, we propose a dual-channel real-time detection framework. Our method jointly leverages head detection and dense optical flow estimation to compute individual velocities, feeding them into a two-branch deep network. We introduce a percentile-based anomaly scoring mechanism coupled with hierarchical clustering to enable interpretable, semantics-aware modeling of motion anomalies—such as stopping or slow walking. The framework adapts dynamically to varying crowd densities, significantly enhancing robustness in anomaly discrimination. Evaluated on multiple dense crowd benchmarks, our approach achieves an 8.2% improvement in mean Average Precision (mAP) and a 22.7% reduction in false positive rate, while maintaining real-time inference speed (>30 FPS). Additionally, it generates physically meaningful velocity-based anomaly heatmaps and category-level explanations, ensuring high accuracy, strong interpretability, and practical deployability.

This work addresses the challenge of limited real infrared data for training deep learning models in PCB defect detection by proposing a cross-modal data augmentation framework that integrates CycleGAN with YOLOv8. The method introduces unsupervised image-to-image translation to synthesize high-fidelity pseudo-infrared images from abundant unpaired visible-light images, preserving both defect structures and thermal distribution characteristics. These generated images are then combined with scarce real infrared samples to train a lightweight detector. Experimental results demonstrate that the proposed approach significantly outperforms baselines trained solely on real infrared data and achieves performance close to fully supervised settings, thereby validating the effectiveness and practicality of cross-modal synthesis for industrial infrared defect inspection.

This work addresses 3D part-level understanding of point clouds without part-level annotations. We propose PointCubeNet, a multimodal framework that partitions point clouds into 3×3×3 voxel blocks and leverages local text-point cloud alignment to generate pseudo-part labels, enabling end-to-end unsupervised training via a designed local contrastive loss. The architecture features dual global-local branches: the global branch captures holistic object semantics, while the local branch performs fine-grained, block-level part reasoning. To our knowledge, this is the first method achieving interpretable and structurally consistent 3D part segmentation and semantic labeling under purely unsupervised settings. Experiments on benchmarks including ShapeNet demonstrate substantial improvements over existing unsupervised approaches. The generated part-level predictions exhibit high geometric plausibility and semantic coherence, significantly enhancing comprehensive 3D object understanding.

Multi-spectral remote sensing imagery poses significant challenges for fine-grained scene understanding in complex terrains (e.g., coastal zones, snow/cloud-covered regions), where conventional RGB-based methods suffer from severe information degradation. To address this, we propose Spectral LLaVA—a novel language-guided framework for multi-spectral scene understanding. It freezes the pre-trained SpectralGPT vision backbone and introduces only lightweight linear projection layers for efficient vision–language alignment. The framework jointly performs multi-spectral encoding, scene classification, and natural language description generation, and is fine-tuned on BigEarthNet v2. Experimental results demonstrate substantial improvements over RGB-only baselines: Spectral LLaVA achieves superior performance in both fine-grained land-cover classification and descriptive text generation—particularly excelling in RGB-degraded scenarios by producing more accurate, semantically richer, and physically grounded captions. This work establishes the first language-guided paradigm for multi-spectral remote sensing interpretation, enabling robust, modality-aware semantic reasoning beyond RGB limitations.

Existing Bayesian motion planning methods are constrained by the light-tailed nature of Gaussian priors, hindering effective exploration of low-probability yet high-quality trajectory regions and thus limiting robotic navigation safety and efficiency. This paper formulates motion planning as a Bayesian inference problem and introduces, for the first time, the Student’s t-distribution as a heavy-tailed prior within a sequential single-sweep smoothing framework—thereby overcoming the exploratory limitations imposed by conventional light-tailed assumptions. We further integrate Monte Carlo sampling with sequential smoothing to generalize the ensemble Kalman smoother. Evaluated in autonomous driving simulations, the proposed method achieves significant improvements in trajectory quality, sampling efficiency, and constraint satisfaction rate. These results empirically validate that heavy-tailed priors substantially enhance the robustness and exploratory capability of probabilistic decision-making in robotics.

This paper investigates whether deterministic computation can be fully characterized by a finite algebraic system analogous to Kleene Algebra with Tests (KAT). Specifically, it addresses whether the deterministic fragment of KAT admits a finite set of control-flow operations sufficient to generate all deterministic finite-state computations. Method: Drawing on formal language theory, algebraic semantics, regular algebra, and model-theoretic techniques, the authors rigorously analyze the expressive limitations of finite algebraic signatures over deterministic automata. Contribution/Results: The paper establishes, for the first time, that no finite set of control-flow primitives can axiomatize the entire class of deterministic finite-state computations—thereby refuting the existence of a KAT-style complete algebraic framework for determinism. This result demonstrates an intrinsic incompleteness in the algebraic characterization of deterministic computation and reveals a fundamental expressiveness boundary of classical control structures (sequence, conditionals, loops), advancing foundational understanding of computational representability in program algebras.

To address the weak, noisy, and high-variance gradient signals inherent in on-policy reinforcement learning for robot policy optimization, this paper proposes a two-stage policy refinement framework: first pretraining an initial policy via Proximal Policy Optimization (PPO), then applying a gradient-free refinement using Triangular-Distribution Evolutionary Strategy (TD-ES). TD-ES employs bounded triangular-distribution perturbations, symmetric sampling, and a center-rank-based finite-difference estimator to ensure bounded exploration while substantially reducing gradient estimation variance. The method is fully gradient-free, highly parallelizable, and computationally lightweight. Evaluated across multiple robotic manipulation tasks, it achieves an average 26.5% improvement in policy success rate over the PPO-only baseline, alongside a marked reduction in training variance. This work establishes a novel paradigm for high-reliability, sample-efficient policy optimization in robotics.