The core capability of artificial intelligence (AI), represented by large language models, stems from statistical modeling over large-scale human experiential data, which can be characterized as a form of machine empiricism. From this perspective, the capacity of AI demonstrates a jagged profile, markedly uneven across dimensions. On this basis, three paradigms for applying AI have emerged: ① synergy, which combines AI with complementary methods to offset its weaknesses; ② AI-native redesign, which leverages emerging capabilities such as perception, reasoning, planning, and reflection to re-engineer tasks as agents; and ③ cross-domain extension, which extends generative modeling beyond natural language to other domains. Furthermore, AI is driving technological equalization by lowering barriers through natural language interfaces, open-source models, and token-based services. This paper concludes by discussing the implications of this perspective for talent cultivation.
Knowledge tracing (KT) aims to infer learners’ dynamic knowledge states from historical interactions and predict their future responses. Despite considerable advances, existing KT methods still suffer from three practical limitations: performance disparities across student groups, sparse interaction data, and insufficient modeling of forgetting dynamics. To address these challenges, this paper proposes HGC-FKT, a heterogeneous counterfactual graph-enhanced fair knowledge tracing model. The model introduces a counterfactual-inspired graph augmentation module, which identifies disadvantaged student groups based on their historical learning records and constructs hypothetical interactions with high-value candidate questions. Potential responses to these unobserved interactions are estimated from the observed behaviors of similar learners and served as augmentation signals to enrich the student-question graph. Meanwhile, a group-disparity regularizer further reduces prediction differences between student groups. To alleviate interaction sparsity, HGC-FKT jointly encodes a student-question graph and a question-knowledge concept graph, thereby integrating behavioral and structural information. Furthermore, a forgetting-aware attention module further incorporates actual time intervals and repeated practice into temporal knowledge-state modeling. Extensive experiments on the Assist09, Assist12, and Junyi datasets demonstrate that HGC-FKT consistently improves predictive performance for disadvantaged students while maintaining strong overall accuracy. The model achieves an AUC of 0.8108 on Assist09, demonstrating the effectiveness of integrating graph augmentation, dual-graph representation learning, and forgetting-aware temporal modeling designs.
Owing to the development of big data and artificial intelligence, industry users are increasingly leveraging machine learning for data analysis. This study focuses on data preparation for machine-learning scenarios to automatically construct a high-quality data-preparation pipeline for a specified machine-learning task (including a tabular dataset, training objective, and model configuration). Existing methods typically rely on deep neural networks and require a predefined and limited operation space for data preparation, making it difficult to satisfy customization demands. Researchers are increasingly investigating the use of large language models (LLMs) to automatically generate data-preparation pipelines. Nevertheless, LLM have difficulty understanding the structure of tabular data and lack performance-oriented optimization for downstream tasks. This ultimately limits model performance. Hence, this study proposes a novel framework named LLMPipe to generate a tabular-data preparation pipeline based on the optimization of large and small models. The method first leverages the advantages of LLM in semantic understanding and knowledge generalization to generate customized and domain-adaptive data-preparation operations by incorporating task descriptions and contextual information. Subsequently, it introduces a lightweight model to model operation features and uses reinforcement learning to optimize the operation ordering strategy, thereby dynamically orchestrating the entire data-preparation pipeline. Based on a collaborative “large-model expansion and small-model tuning” approach, the framework preserves the ability to understand domain knowledge and compose high-quality pipelines while improving overall computational efficiency and effectively enhancing the customizability and rationality of the generated pipelines. Extensive experiments are conducted across multiple real-world machine-learning tasks from Kaggle. The results reveal that the proposed method significantly outperforms existing methods, with state-of-the-art accuracy achieved across machine-learning tasks.
Multimodal interaction has become an important research direction in artificial intelligence. The audio modality carries rich temporal dynamic information and environmental semantics, playing a significant role in scenarios such as intelligent assistants, virtual digital humans, and autonomous driving. However, most existing surveys focus on the visual modality and lack a systematic review of audio-centric cross-modal interaction, making it difficult to comprehensively present the technical roadmap and developmental bottlenecks in this field. Taking audio as the dominant modality, this paper focuses on three major interaction lines—audio-text, audio-static image, and audio-dynamic video—and extends them to audio-driven three-dimensional motion generation. It systematically summarizes representative methods, technical routes, and application progress developed in recent years. On this basis, the paper analyzes the current development status and main limitations of different research directions from the perspectives of data foundation, time-series modeling, evaluation systems, and adaptability to real scenarios, with particular emphasis on the bottlenecks faced by multimodal audio datasets in terms of scale, annotation quality, and acquisition cost. Overall, audio-driven cross-modal interaction has achieved significant progress in unified representation learning, generative modeling, and real-scenario applications, and continues to advance toward stronger multi-granularity alignment capabilities, higher-quality data support, more robust evaluation systems, and more efficient deployment methods.
The development of multimodal large models has transformed visualization charts from static information carriers into important objects for model understanding, reasoning, generation, and safety evaluation. This review summarizes recent advances in visualization and visual analytics in the era of large models, focusing on two major directions: visualization chart understanding and misleading visualization analysis. The former investigates how models recover chart structures, data relations, and visual encodings from different input forms, including bitmap images, vector graphics, and D3 charts, and further supports chart question answering, visual grounding, style design, and the generation of titles and descriptions. The latter examines how misleading charts emerge from data selection, visual encoding, textual explanation, and communication context, with emphasis on taxonomy, evaluation of visual language models, automatic detection, explainable analysis, interactive correction, and safety defense. Existing studies indicate that both chart understanding and misleading visualization analysis rely on reliable alignment among data, visual structures, text, and semantics. Future research should develop unified chart representations, fine-grained evidence grounding, trustworthy generation, human-AI collaborative correction, and visualization defense mechanisms to enhance the safety of large models.
In bulk industrial logistics, order text typically contains implicit compatibility constraints that are challenging to identify accurately; disregarding destination spatial distribution during order splitting causes cross-region detours and exacerbates long-tail order accumulation. This paper proposes a collaborative scheduling framework integrating semantic awareness and uncertainty quantification for steel logistics. The semantic awareness module employs a domain-adapted sentence embedding model based on Bidirectional Encoder Representations from Transformers, latent relation graphs, and a mixture-of-experts routing module to model material compatibility, with industrial rules incorporated via large language model distillation for zero-shot inference on long-tail orders. The decision module adopts a variational hypernetwork that generates the mean and variance of network weights via variational inference, thus explicitly quantifying epistemic uncertainty and mitigating over-optimistic estimates for rare materials. Discrete routes and continuous split ratios are jointly optimized under a parameterized Markov decision process, with a differentiable physical constraint layer enforcing weight limits. On 1.7 million real industrial records, the framework achieved a competitive ratio of 0.8641 at approximately 100000 orders and a precision of 0.8942 in long-tail settings.
中文核心期刊
