Embodied robots need more than a description of the current image: they must estimate how the scene may change under motion, contact, and partial observation. Vision-centric world models provide this predictive layer, but they expose it through different state interfaces. This paper organizes the literature into four families according to the state available to planning: future observations, compact latent states, geometry-structured maps, and persistent entities or relations. The comparison examines how each state is formed, advanced, and queried; where inference and planning costs arise; and which errors matter in physical use. Representative methods show that observation prediction retains interpretable appearance but makes repeated rollout expensive. Latent dynamics reduce that cost while risking the loss of contact-scale variables. Geometric states support pose, clearance, and occupancy queries, although their reliability depends on calibration and timely updates. Entity-relational states preserve object identity and task relations, yet they remain vulnerable to binding errors under occlusion. Evaluation is therefore linked to the exposed state rather than to a single generic score. The resulting framework clarifies where temporal prediction, persistent geometry, and semantic identity complement one another in embodied planning.

Show more

Museum digital transformation is shifting exhibitions from static display toward data-driven cultural experience design. However, many intelligent guide systems still emphasize retrieval, navigation, and efficiency rather than narrative relationships, aesthetic rhythm, and visitor understanding. This study proposes an AI-driven framework for narrative generation and visitor path optimization using open collection metadata, gallery spatial graphs, and simulated visitor behavior. Collection records from 2020 to 2025 are gathered from Smithsonian Open Access, Cleveland Museum of Art, Europeana, and Tate. Knowledge graphs, BERT-based semantic encoding, and large language models are used to generate cultural narrative chains, while Dijkstra's algorithm and NSGA-II optimize visitor routes. Results show that the AI Curatorial Constraint strategy outperforms chronological and semantic-clustering approaches in thematic coverage, coherence, and curatorial consistency. Narrative-optimized routes also improve continuity and reduce congestion while maintaining low movement cost. The framework supports intelligent curation, cultural communication, and public aesthetic education.

Show more

From the perspective of the current development trend of large language models, full evaluation on complete test sets entails high computational, time, and financial costs and cannot meet the need for frequent and rapid testing. Anchor-point evaluation relies on a small number of representative anchor samples to approximately assess the overall performance of a model, making it a feasible solution for low-cost and efficient evaluation. However, this method depends heavily on the quality of the anchor samples and the design of the evaluation framework, and it has evident shortcomings in terms of data distribution shift, sample noise, experimental conditions, and error control. This paper summarizes various problems in anchor-point evaluation from the two levels of datasets and model evaluation, compares optimization approaches such as active sampling, domain adaptation, data denoising, weighted evaluation, and multidimensional robust evaluation, and discusses the scope of application and constraints of general optimization solutions. Active sampling and lightweight data correction can improve the stability and accuracy of anchor-sample evaluation; in contrast, complex domain adaptation, multilevel adversarial experiments, and large-scale error analysis have excessively high computational costs, weakening the core lightweight advantage of anchor-point evaluation. Future research needs to balance evaluation accuracy, computational efficiency, and actual business needs, with an emphasis on developing lightweight and reliable anchor-point evaluation solutions suited to few-sample scenarios.

Show more

The large pre-trained models centered on the Transformer architecture have been widely applied in fields such as government affairs, healthcare, and cultural creation, significantly enhancing the efficiency of content production. However, they have also given rise to multiple security and ethical risks: Data leakage can lead to the leakage of users' medical records privacy, attackers can steal commercial large models through API calls, multi-modal deep fakes are used in telecommunications fraud, and algorithmic biases in similar Amazon recruitment large models can cause employment discrimination. This paper takes generative large models as the research object and adopts methods such as literature review, case analysis, and comparative research to classify the risks of large models into two categories: information security and ethical security. It summarizes the protection technologies throughout the entire life cycle of training, reasoning, and traceability, compares the differences in AI regulatory systems between China and abroad, and constructs a complete governance path from the technical, institutional, and industrial dimensions. This paper can provide a basic theoretical reference for the safe implementation of large model enterprises and the improvement of industry norms by regulatory authorities.

Show more