Background: Text-to-Structured Query Language (SQL) systems, which allow users to query databases using natural language, have advanced significantly with the rise of Large Language Models (LLMs). While this progress has boosted accuracy, it has also introduced serious security and practical deployment challenges that existing literature has not systematically analyzed. Methodology: In this article, we systematically review the existing literature to examine various approaches, identify security vulnerabilities, discuss design trade-offs, and outline future challenges. Our goal is to offer a comprehensive overview. Results: Our analysis shows that modern approaches fall into two main categories: prompt engineering on proprietary models and fine-tuning open-source models. Regarding security, using the Open Worldwide Application Security Project (OWASP) Top 10 framework, we identify critical threats such as Prompt Injection (P2SQL), data poisoning to create backdoors, and inference attacks. The analysis reveals that current defense mechanisms are not effective enough against these attacks. We also highlight the strategic trade-off between the superior accuracy of proprietary models and the control, security, and cost-effectiveness of open-source models. Conclusion: This survey provides a systematic analysis of security vulnerabilities in LLM-based Text-to-SQL systems, concluding that current countermeasures are inadequate. Our findings point to an urgent research direction: developing systems that are not only accurate but also robust, efficient, and fundamentally secure for reliable real-world deployment.

Geometric morphometrics is a foundational tool for quantifying biological shape variation, but the manual digitization of landmarks remains a major bottleneck, especially for small or rare-species datasets. Existing automated methods based on handcrafted features can operate with few annotated samples but often require careful descriptor design and parameter tuning, whereas modern deep learning approaches typically need hundreds to thousands of annotated images. In this paper, we propose a three-stage deep learning framework that combines self-supervised representation learning with few-shot heatmap-based fine-tuning for automatic biological landmark detection, with a primary focus on insect-wing morphometrics but extending to other two-dimensional biological structures. In the first stage, we pretrain a Masked Autoencoder (MAE) on a pool of unlabeled biological images, optionally starting from public ImageNet weights to leverage natural-image priors. In the second stage, the pretrained encoder is adapted separately to each target dataset using a lightweight heatmap-regression head and only a small number of annotated images. The same pipeline is applied without dataset-specific architectural changes across nine public benchmarks covering diverse subjects, resolutions, and numbers of landmarks per specimen, demonstrating the versatility of our method. The framework reaches near-converged performance with 7 to 20 annotated images per dataset, improves over the best previously reported baseline on six of the nine benchmarks, and remains competitive with data-intensive deep learning methods at a fraction of their annotation cost. The source code is released as open source to facilitate reproducibility and adoption by the biological morphometrics community.

Vision–Language–Action (VLA) models aim to unify perception, language grounding, and robot control, yet recent progress in robotic manipulation remains difficult to compare because systems differ widely in observation modalities, action interfaces, execution regimes, control horizons, and measurement boundaries. The counted analytical corpus comprises 59 unique works: 20 VLA systems in the settings-tuple audit, eight benchmark or evaluation-protocol studies, and 31 additional mechanism or context studies. Survey seeds and general background references are not included in this count. We organize the field through a five-axis design-space taxonomy: (A) perception encoding and spatial grounding; (B) action generation, control regime, and reasoning–control interface; (C) training and adaptation; (D) data provenance and embodiment; and (E) temporal horizon and stabilization. Across these axes, we synthesize five recurring reported limitation patterns—visual brittleness, action degeneration, static supervised fine-tuning limitation, embodiment gaps, and long-horizon breakdown—and consolidate them into four research gaps: reasoning–action decoupling, closed-loop control under contact, spatiotemporal grounding, and data/embodiment mismatch. To support auditable cross-paper interpretation, we introduce a comparability-first reporting schema, a taxonomy coding manual, canonical pseudocode for core VLA mechanisms, and an evidence-tier policy for calibrating claims. We further critique what LIBERO, SimplerEnv, and CALVIN can and cannot establish, showing that current benchmarks remain insufficient for strong deployment-level claims about contact-rich recovery, viewpoint robustness, embodiment transfer, long-horizon stability, and end-to-end latency.