Agents-K1: Building an Agent-native Scientific Knowledge Graph System

This article introduces Agents-K1—an end-to-end knowledge orchestration pipeline designed to convert raw scientific documents into agent-native knowledge graphs. The system addresses the shortcomings of current LLM-based research agents in scientific knowledge orchestration, with core components including a multimodal parser, a 4-billion-parameter information extraction model, and a unified retrieval interface. Based on this, it has built Scholar-KG, a large-scale knowledge graph covering 2.46 million cross-disciplinary papers, laying the foundation for the scientific reasoning capabilities of intelligent agents. Source: Published on arXiv in June 2026, original title: Agents-K1: Towards Agent-native Knowledge Orchestration.

Current LLM-based research agents have made significant progress in task orchestration, but there are obvious gaps in scientific knowledge orchestration: existing methods often simplify academic papers into abstracts, surface mentions, and flat citation relationships, ignoring entities, claims, evidence, mechanisms, and method lineages necessary for scientific reasoning. This prevents agents from understanding the deep structure of literature (such as conceptual causal relationships, strength of evidence support, and methodological inheritance), severely limiting their utility in scenarios like research assistance and literature review.

The Agents-K1 system architecture integrates three core components:

1. Multimodal Parser: Processes full paper texts, including five modules: entity recognition, multimodal evidence extraction, citation relationship analysis, inter-entity relationship recognition, and cross-document alignment;
2. 4B-parameter Information Extraction Model: Trained using the GRPO algorithm, combining rule-based reward mechanisms and multi-task learning to improve extraction quality and reasoning efficiency;
3. GraphAnything CLI Unified Retrieval Interface: Supports web search, multimodal graph retrieval, and cross-document traversal to enable interactive knowledge exploration.

Scholar-KG, built based on the Agents-K1 pipeline, has the following features:

• Scale: Covers 2.46 million cross-disciplinary papers across 6 fields including computer science, physics, and biology; a 1 million subset is publicly available;
• Features: Fine-grained entity representation (including methods, datasets, etc.), multimodal evidence association, dynamic update mechanism, and explainable reasoning paths (each claim traces back to original evidence).

Experimental evaluations show that Agents-K1 outperforms others:

• Metrics: Excellent performance in entity recognition accuracy, F1 score for complex relation extraction, evidence location precision, and cross-document alignment;
• Comparison: Full-text parsing provides significant information gain compared to abstract-only methods; fine-grained relation extraction improves graph connectivity; multimodal processing effectively utilizes non-text information.

Agents-K1 has good scalability and application prospects:

• Scalability: Supports general domain adaptation, user-defined extraction patterns, and synthetic data generation;
• Potential Scenarios: Intelligent literature review, research trend analysis, cross-disciplinary knowledge discovery, and research assistants (precision retrieval and evidence support).

Current limitations and future directions:

• Limitations: Mainly processes English literature, has knowledge update delays, and limited complex multi-hop reasoning capabilities;
• Future: Introduce stronger base models to improve extraction quality, explore neuro-symbolic integration to enhance reasoning interpretability, and develop interactive knowledge completion mechanisms combining human feedback to improve the graph.

Agents-K1 represents an important progress in the knowledgeization of scientific literature. Through fine-grained multimodal parsing, efficient extraction models, and a unified retrieval interface, it provides scientific knowledge infrastructure for intelligent agents. The openness of Scholar-KG is expected to promote the practical application of intelligent agents in research assistance, education, decision support, and other fields, marking a key shift from 'agent orchestration' to 'knowledge orchestration'.