Zongsheng Cao, Bihao Zhan, Jinxin Shi, Jiong Wang, Fangchen Yu, Zhijie Zhong, Zijie Guo, Tianshuo Peng et al.
An end-to-end pipeline that converts scientific papers into structured knowledge graphs directly usable by agents.
Current LLM-based research agents focus on agent orchestration but overlook scientific knowledge orchestration. Papers are reduced to abstracts, surface mentions, and simple citation edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning.
Agents-K1 integrates three components: (1) a multimodal parser with a five-module schema that extracts entities, multimodal evidence, citations, and typed inter-entity relations from the full paper; (2) a 4B information extraction backbone trained with GRPO under a rule-based reward; and (3) a graph-anything CLI, a tri-source agent interface unifying web search, multimodal graph retrieval, and cross-document traversal. This pipeline processes 2.46 million scientific papers to produce Scholar-KG, with a one-million-paper subset released.
Achieves superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning. The same pipeline can be extended to general-domain corpora and schema-conformant data synthesis.