To map the knowledge structure of large language model (LLM) applications in the dental field (LADF) through a dual-database scientometric study that highlights trends, collaborations, hotspots, and future directions. The Web of Science and Scopus were searched on August 1, 2025. LLM-related articles in dentistry were screened. After removing duplicates, 311 English articles and reviews were included in the study. Bibliometrix (v5.0) was used for characteristic and citation analyses. CiteSpace (v6.4.R1) was used for keyword analysis. The number of publications surged from 24 (2023) to 182 (Jan–Aug 2025). Among all publications, 90% were original research articles. Two-thirds were open-access, funded mainly by governments. As for the collaboration network, author and institutional networks were fragmented. National-level collaboration was stronger. High-income countries dominated the LADF output. Keyword clustering revealed a hub-and-spoke structure that included three research frontiers: 1. broadening applications, 2. deepening clinical use, and 3. comparative evaluation, with accuracy and quality as central themes. LADF has expanded rapidly, but research remains fragmented. Shared datasets, stronger global collaboration, and the development of standardized evaluation metrics, particularly for diagnostic and question-answering proficiencies, are necessary to address the central themes of accuracy and quality.

Regulatory approval may occur before or after a research topic becomes prominent in the scientific literature. We examined this temporal relationship using botulinum neurotoxin (BoNT) as a longitudinal case study. We analyzed 2065 anatomy-related BoNT publications indexed in Web of Science and Scopus from 1983 to 2024. The literature was divided into 16 periods based on U.S. FDA approval dates for onabotulinumtoxinA, and indication-related keyword rankings and thematic clusters were compared with the corresponding approval dates. Among the 10 indication groups examined, 7 showed prominent indication-related keywords before FDA approval, 2 showed signals around the time of approval, and 1 showed a post-approval signal. We describe these patterns as “proactive exploration” when scientific attention became prominent before approval and “reactive catalysis” when it became more prominent around or after approval. Differences in the observed lag may be related to target anatomical complexity and market orientation, but this remains conceptual. Overall, the findings show that time-resolved scientometrics can retrospectively identify scientific patterns associated with regulatory milestones. The approach is hypothesis-generating and requires prospective validation before it can be used for regulatory forecasting.