[2608.29612v1] LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge

The paper "LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge" (arXiv:2608.29612v1) introduces ASKS (Agent-Driven Scientific Knowledge System), a framework for scientific knowledge compilation that preserves source evidence while organizing scientific data.

The system operates through three distinct computational roles: LLMs Interpret: Large Language Models convert raw scientific sources into human-readable Wiki views and machine-facing semantic slots. Embeddings Organize: Semantic embeddings place local and accumulated knowledge into a shared semantic geometry, using similarity signals to guide identity, routing, and membership decisions. * Graphs Emerge: Deterministic rules and validated GraphDeltas integrate these semantic insights into a persistent, source-linked knowledge graph that evolves as new papers are ingested.

Validated through a longitudinal demonstration compiling 56 papers from a single research program, ASKS demonstrates that this approach creates a stable, low-churn research map with clear lineage and cross-paper support, effectively bridging the gap between flexible LLM interpretation and rigorous scientific authority.

The material describes an agent-driven pipeline for compiling scientific literature into structured knowledge representations by combining three complementary layers: LLM-based interpretation, embedding-based organization, and graph-based synthesis. Rather than treating language models as a single black-box summarizer, the approach positions them as interpreters that extract claims, concepts, methods, and relationships from papers; embeddings then provide a continuous semantic space for grouping related content, identifying overlaps, and organizing large corpora; and knowledge graphs emerge as an explicit, inspectable structure over the extracted entities and relations. The central insight is that scientific knowledge compilation benefits from separating understanding, organization, and representation: LLMs handle local meaning and context, embeddings handle global similarity and clustering, and graphs capture durable relational structure that can support downstream reasoning and retrieval.

Its key contribution is a scalable architecture for automated scientific knowledge compilation that moves beyond flat document stores or one-off LLM summaries toward a persistent, queryable knowledge graph. By making the graph “emerge” from interpreted and embedded content, the pipeline can potentially reduce manual curation, expose cross-paper connections, and support tasks such as literature synthesis, method comparison, gap discovery, and provenance-aware reasoning. This matters because scientific corpora are large, heterogeneous, and rapidly growing; a system that can continuously interpret new papers, embed them into an evolving semantic space, and update an explicit knowledge graph offers a practical path toward maintainable, machine-actionable scientific knowledge bases rather than ephemeral natural-language summaries.