Multidimensional Knowledge Profiling Achieves Insights from 100,000 Scientific Papers

Scientists are tackling the challenge of synthesising the ever-growing body of research in fields like machine learning and artificial intelligence. Zhucun Xue, Jiangning Zhang, and Juntao Jiang from Zhejiang University, alongside colleagues including Jinzhuo Liu, Haoyang He, and Teng Hu, present a novel approach to understanding scientific progress through large-scale knowledge profiling. Their work addresses a critical limitation of current bibliometric tools, which often lack detailed semantic analysis, by constructing a comprehensive database of over 100,000 papers from leading conferences between 2020 and 2025. This research is significant because it provides an evidence-based view of evolving research themes , revealing shifts towards areas like AI safety and agent-oriented studies, and offering valuable insights into methodological transitions and emerging trends for the wider scientific community.

Scientists find current research tools rely mainly on metadata and offer limited visibility into the semantic content of papers, making it hard to track how research themes evolve over time or how different areas influence one another. To obtain a clearer picture of recent developments, researchers compiled a unified corpus of more than 100,000 papers from 22 major conferences between 2020 and 2025. They construct a multidimensional profiling pipeline to organise and analyse their textual content. By combining topic clustering, LLM-assisted parsing, and structured retrieval, they derive a comprehensive representation of research activity that supports the study of topic lifecycles, methodological transitions, dataset and model usage.

AI Research Landscape Mapping with LLMs

Scientists are increasingly challenged by the scale and diversity of contemporary AI research. Across computer vision, machine learning, natural language processing, and related areas, the past five years have seen rapid shifts in model architectures, training strategies, datasets, benchmarks, and application domains. This expansion makes it difficult to situate individual works within broader developments, track how research themes evolve, or identify emerging, stabilizing, or declining areas. Traditional bibliometric methods, built on metadata, co-citation networks, and keyword statistics, provide high-level overviews but capture limited semantic information and treat topics as largely static entities.

Recent systems incorporating large language models (LLMs) demonstrate improved semantic analysis, supporting tasks such as retrieval-augmented question answering and automated survey generation. However, these tools are typically designed for short-range retrieval, single papers, or narrow tasks, and do not provide a coherent, longitudinal view of large scientific corpora. These gaps highlight the need for a unified way to organize, summarize, and interpret the rapidly expanding body of AI literature. In this work, researchers construct a large-scale profiling pipeline aimed at characterizing the recent development of AI research.

Using more than 100,000 papers from 22 major conferences published between 2020 and 2025, they combine text clustering, LLM-assisted semantic parsing, and lightweight retrieval techniques to form a structured representation of research problems, methods, datasets, and topical dynamics. Rather than emphasizing algorithmic novelty, the focus is on creating a coherent analytic framework that enables researchers to explore and reason about the field at multiple levels of granularity. This study provides two complementary benefits. First, it derives a high-resolution view of topic lifecycles, dataset and model adoption patterns, and methodological transitions across areas such as vision, multimodal learning, foundation models, and generative modeling.

Second, by incorporating structured retrieval and semantic filtering, it enables grounded, evidence-based queries that support practical research tasks, such as surveying subfields, tracing the evolution of techniques, or identifying emerging directions. Through this analysis, they highlight several notable shifts in the AI landscape, including the consolidation of previously fast-moving areas, the rise of multimodal and agent-oriented research, and clear transitions in compute usage and model scaling practices. The resulting knowledge database is expected to serve as a resource for understanding broad trends, informing future meta-analyses, and supporting data-driven research planning. The work makes the following contributions: construction of a large-scale profiling pipeline over more than 100,000 papers, integration of clustering-based topic organization with LLM-assisted parsing and retrieval, and comprehensive empirical analyses revealing trends in topic evolution, emerging subfields, methodological transitions, dataset and model dynamics, and institutional research patterns.

Overall, these results provide an evidence-based view of how modern AI research is evolving and offer a foundation for transparent, large-scale, and semantically grounded scientometric analysis. Content mining and trend analysis of scientific literature are important topics in scientometrics and information science. Traditional bibliometric approaches rely on metadata, extracting abstracts, authors, journals, keywords, and citations, and applying co-citation and co-occurrence analysis. This framework relies on semantic understanding rather than mere keywords, enabling the discovery of emerging or previously underexplored research topics, and finally obtaining information such as metadata, core content summaries, technical details, analysis of novelty, and system requirements.

Large language models have shown potential for literature retrieval and summarization, but direct application in research workflows faces key limitations: hallucination, omission or misinterpretation of complex information, and poor grounding in evidence, particularly over large, multi-year corpora. To address these challenges, an intent-driven hierarchical retrieval pipeline is developed over the structured ResearchDB, combining metadata filtering with weighted multi-field semantic search to provide reliable, evidence-based input to language models. To handle complex queries, the query is first decomposed into simpler sub-questions that isolate distinct information needs. Each sub-question is then processed using semantic parsing to generate structured retrieval instructions in JSON format, specifying relevant keywords, entities, and content types for subsequent retrieval.

Large-scale analysis of machine learning research trends reveals

Scientists have compiled a unified corpus of over 100,000 papers from 22 major conferences spanning 2020 to 2025, enabling a detailed analysis of research trends in machine learning, vision, and language. The team constructed a multidimensional profiling pipeline to organise and analyse the textual content of these publications, moving beyond traditional bibliometric tools reliant on metadata alone. By integrating topic clustering, large language model (LLM)-assisted parsing, and structured retrieval, researchers derived a comprehensive representation of research activity, facilitating the study of topic lifecycles and methodological transitions. Experiments revealed the successful extraction of crucial information from each paper, including metadata such as title, authors, conference, year, and citation counts, alongside semantic content like research problems, keywords, and contributions.

The work employed minerU to convert PDF files into structured Markdown, then leveraged Deepseek-R1-32B to perform multi-dimensional analysis, focusing on aspects like research questions, methods, datasets, and limitations. Topic clustering, utilising UMAP for dimensionality reduction and HDBSCAN, resulted in the distinction of over 300 topic categories based on semantic similarity, demonstrating a granular level of analysis. The analysis highlighted notable shifts, including growth in safety, reasoning, and agent-oriented studies, alongside the stabilisation of areas like neural machine translation and graph-based methods. The pipeline’s ability to systematically extract and organise semantic information provides a dynamic, hierarchical view of knowledge evolution, enabling trend analysis and cross-domain comparisons. This breakthrough delivers a resource for understanding broader trends and identifying emerging directions in the rapidly evolving field of artificial intelligence, offering a foundation for data-driven exploration and research planning. The resulting ResearchDB supports fine-grained topic investigation and facilitates evidence-based research decision-making.

AI Research Landscape Mapped by Framework Reveals Key

Scientists have developed a multidimensional knowledge profiling framework to analyse over 100,000 artificial intelligence papers published between 2020 and 2025. This framework combines topic clustering, large language model-based semantic parsing, and hierarchical retrieval to map research dynamics, dataset and model trends, and institutional patterns. The resulting system enables evidence-driven discovery of emerging research directions, methodological shifts, and evolving technical paradigms within the field. The analysis highlights notable shifts, including growth in safety, reasoning, and agent-oriented studies, alongside the stabilisation of areas like neural machine translation and graph-based methods. By integrating semantic understanding with structured retrieval, the framework offers both a broad overview of the research landscape and detailed, topic-level perspectives, supporting trend identification and research planning.