A new AI pipeline treats scientific papers as structured knowledge, not just text to summarize.
Researchers introduced Agents-K1, an end-to-end system that converts raw academic documents into what they call "agent-native" scientific knowledge graphs. Rather than reducing papers to abstracts and flat citation links, the pipeline extracts entities, claims, evidence, mechanisms, and method lineages from the full text. It uses a multimodal parser with a five-module schema, a 4-billion-parameter information-extraction model trained with a rule-based reward method called GRPO, and a command-line interface that combines web search, graph retrieval, and cross-document traversal. The team used it to process 2.46 million scientific papers across six subject areas, producing a knowledge graph called Scholar-KG, with a one-million-paper subset now publicly released.
Most AI research tools treat citations as edges on a flat graph and abstracts as the unit of meaning — which is a bit like summarizing a book by reading only its table of contents. Agents-K1 argues the field has been optimizing agent orchestration while neglecting what those agents actually know. If the approach holds up under broader use, it could meaningfully change how AI systems assist with literature review and hypothesis generation.
The system is also designed to extend to non-scientific domains, which is either a sign of genuine generality or a hedge against a niche use case — the paper doesn't make that case in detail.