Agents-K1: Towards Agent-native Knowledge Orchestra...

Agents-K1: Towards Agent-native Knowledge Orchestra... | AI Research

Key Takeaways

Agents-K1: Towards Agent-native Knowledge Orchestration is a framework designed to transform raw scientific papers into structured, "agent-native" knowledge...
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 \texttt{cites} edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning.
To this end, we introduce \textbf{Agents-K1}, an end-to-end knowledge orchestration pipeline that converts raw documents into agent-native scientific knowledge graphs.
On top of this, we process 2.46 million scientific papers across six subjects to produce \textbf{Scholar-KG}, of which we release a one-million-paper subset, and the full Scholar-KG is accessible via the SCP link below.

Paper AbstractExpand

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 \texttt{cites} edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning. To this end, we introduce \textbf{Agents-K1}, an end-to-end knowledge orchestration pipeline that converts raw documents into agent-native scientific knowledge graphs. Agents-K1 integrates three components under a unifying theoretical foundation: a multimodal parser whose five-module schema captures entities, multimodal evidence, citations, and typed inter-entity relations across the full paper rather than abstracts alone; a 4B information-extraction backbone trained with GRPO under a rule-based reward; and a graphanything CLI, a tri-source agent interface that unifies web search, multimodal graph retrieval, and cross-document traversal. On top of this, we process 2.46 million scientific papers across six subjects to produce \textbf{Scholar-KG}, of which we release a one-million-paper subset, and the full Scholar-KG is accessible via the SCP link below. The same pipeline can be extended to general-domain corpora and to schema-conformant data synthesis. Extensive experiments demonstrate that Agents-K1 achieves superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning.

Agents-K1: Towards Agent-native Knowledge Orchestration is a framework designed to transform raw scientific papers into structured, "agent-native" knowledge graphs. While existing research agents often rely on simple summaries or flat citation lists, this project aims to provide a more robust infrastructure that captures the full depth of scientific literature—including figures, tables, equations, and complex relationships between claims and methods. By doing so, it enables AI research agents to perform more reliable, traceable, and evidence-based reasoning.

Moving Beyond Abstracts

Current research agents often struggle because they treat scientific papers as simple text files, frequently ignoring the critical data buried in non-textual elements. Agents-K1 addresses this by using a multimodal parser that treats text, figures, tables, and equations as interconnected evidence. Instead of just noting that one paper cites another, the system identifies the specific intent behind that citation—such as whether a paper is extending a method, challenging a claim, or simply using a baseline. This creates a rich, navigable map of scientific knowledge rather than a collection of isolated documents.

A Three-Part Pipeline

The framework is built on three core components that work together to turn raw documents into actionable intelligence:

Multimodal Parsing: The system decomposes papers into content units and uses "semantic anchors" to bridge different types of information. This allows the agent to connect a specific table or equation to the textual claims that support or explain it.
Extraction Backbone: The team developed a 4B-parameter model trained using reinforcement learning (GRPO) to ensure that the extracted data is accurate, valid, and structured. This model is designed to be efficient and adaptable to different domains.
GraphAnything CLI: This interface acts as the "bridge" for AI agents. It unifies web search, graph retrieval, and cross-document traversal, allowing an agent to move through the knowledge graph to verify facts, compare research methods, and synthesize new ideas.

Large-Scale Knowledge Construction

To demonstrate the power of this approach, the authors processed 2.46 million scientific papers across six disciplines, including computer science, biology, and physics. This resulted in the creation of "Scholar-KG," a massive knowledge graph that organizes entities, claims, and method lineages. A one-million-paper subset of this data has been released for community research, and the underlying pipeline is flexible enough to be applied to other document collections beyond scientific literature.

Impact on Research Reasoning

Experiments show that Agents-K1 significantly improves the performance of research agents. By providing a structured, verifiable knowledge base, the system helps agents avoid the "blind spots" common in text-only retrieval. For example, the framework demonstrated substantial gains in accuracy for complex multi-hop reasoning tasks and research-oriented benchmarks, proving that when agents have access to a well-organized map of scientific evidence, they can produce more grounded and reliable research outcomes.