Agents-K1: Towards Agent-native Knowledge Orchestration

A new pipeline called Agents-K1 converts raw scientific documents into structured knowledge graphs, moving beyond abstract-level summaries to capture entities, evidence, and method lineages across full papers, according to research posted on arXiv [1]. The system processes entire papers through a multimodal parser with a five-module schema, a 4B-parameter information-extraction model trained with group relative policy optimization under a rule-based reward, and a command-line interface called graphanything that unifies web search, multimodal graph retrieval, and cross-document traversal [1][2]. The authors argue that existing research agents reduce papers to abstracts and flat citation edges, omitting claims, mechanisms, and evidence chains that scientific reasoning requires [2]. To demonstrate the pipeline, the team processed 2.46 million scientific papers across six subjects, producing a resource named Scholar-KG [1][2]. A one-million-paper subset has been released, while the full graph is accessible through a provided SCP link [1]. The same architecture can be extended to general-domain corpora and schema-conformant data synthesis, the paper states [2]. Experiments reported in the preprint indicate that Agents-K1 achieves superior performance on scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning tasks [1][2]. The work appears on arXiv under the Computer Science and Artificial Intelligence category, submitted on 11 June 2026, and is hosted through arXivLabs, a framework that allows collaborators to develop and share new features on the platform [1]. The release comes amid broader efforts to structure scientific knowledge computationally. Knowledge graphs have been applied across domains ranging from biomedical research to sustainability monitoring. The United Nations Sustainable Development Goals, adopted in 2015, rely on tracking cross-cutting indicators across 17 goals, a task that structured knowledge representations could support [7]. Similarly, molecular biology has long depended on structured databases to catalog entities such as transcription factors, the roughly 1,600 proteins in the human genome that regulate gene expression by binding to specific DNA sequences [8]. Agents-K1 extends this principle to the full text of scientific literature, aiming to preserve the relational richness that abstracts alone discard [2].