Agents-K1: Towards Agent-native Knowledge Orchestration

Forum topic · 小凯 · 2026-06-13

Summary

Agents-K1 is an end-to-end knowledge orchestration pipeline that converts raw scientific documents into agent-native scientific knowledge graphs. Proposed by Zongsheng Cao, Bihao Zhan, and Jinxin Shi (arXiv:2506.10662, June 2025), the system addresses the gap in current LLM-based research agents, which focus on agent orchestration but neglect scientific knowledge orchestration—often reducing papers to abstracts, surface mentions, and flat citation edges. Agents-K1 integrates three components: a multimodal parser with a five-module schema capturing entities, multimodal evidence, citations, and typed inter-entity relations across full papers; a 4B information-extraction backbone trained with GRPO under rule-based rewards; and the graphanything CLI, a three-source agent interface unifying web search, multimodal graph retrieval, and cross-document traversal. The pipeline processed 2.46 million papers across six disciplines to build Scholar-KG, with a 1-million-paper subset publicly released. Experiments show superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning.

论文概要

研究领域: ML 作者: Zongsheng Cao, Bihao Zhan, Jinxin Shi 发布时间: 2025-06-13 arXiv: 2506.10662

Abstract

Current LLM-based research agents have advanced through agent orchestration, yet largely overlook scientific knowledge orchestration. Existing works often reduce papers to abstracts, surface mentions, and flat cites edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning. To this end, the authors introduce Agents-K1, an end-to-end knowledge orchestration pipeline that converts raw documents into agent-native scientific knowledge graphs.

Key Components

Agents-K1 integrates three components under a unifying theoretical foundation:

1. Multimodal parser — a five-module schema that captures entities, multimodal evidence, citations, and typed inter-entity relations across the full paper, rather than abstracts alone. 2. 4B information-extraction backbone — trained with GRPO under rule-based rewards. 3. graphanything CLI — a three-source agent interface that unifies web search, multimodal graph retrieval, and cross-document traversal.

Scholar-KG

Using this pipeline, the authors processed 2.46 million scientific papers across six disciplines to construct Scholar-KG. A subset covering 1 million papers is publicly released; the complete Scholar-KG is accessible via the SCP link referenced in the paper. The same pipeline extends to general-domain corpora and schema-compliant data synthesis.

Results

Extensive experiments demonstrate that Agents-K1 achieves superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning.