Title: Bridging AI Research and the Humanities: A Multi-Agent Approach to Evidence-Based Scholarship

Abstract:

While large language model (LLM)-driven research assistants have made significant strides in scientific and engineering disciplines—fields characterized by executable experiments, code, and quantitative metrics—the humanities demand a distinct intellectual approach. Humanistic inquiry relies on interpretive reasoning, constructing arguments grounded in evidence from primary sources, where academic merit hinges on accurate quotation, traceable provenance, and meticulous close reading. Current research tools are predominantly designed for task execution and information retrieval, lacking the capacity for evidence-based interpretive analysis.

To bridge this divide, we present SPIRE (Scholarly-Primitives-Inspired Research Engine), a multi-agent framework tailored for humanities scholarship. Inspired by Scholarly Primitives theory, SPIRE transforms common humanities tasks into collaborative agent roles, including source discovery, evidence annotation, comparative analysis, provenance verification, sampling, citation binding, and argumentative synthesis. These agents operate on a multi-scale close-reading infrastructure comprising individual passages, intra-context graph communities, and cross-context semantic clusters.

Evaluated on a benchmark of peer-reviewed papers focused on classical Chinese and Greco-Roman Latin studies, SPIRE demonstrated superior reliability in recovering cited primary-source evidence compared to Naive LLM, Text RAG, and GraphRAG models. Additionally, it achieved higher scores in blind evaluations across metrics such as answer accuracy, depth, coverage, and evidence quality. Ablation studies confirm that both the specialized scholarly-operation agents and the close-reading retrieval mechanisms are essential for producing evidence-grounded essays. The framework’s code, data catalogs, and reproduction scripts are publicly available at https://github.com/YatingPan/SPIRE.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC