Problem

The rapid expansion of financial research, especially in option pricing, makes traditional systematic literature reviews (SLRs) infeasible because manual screening and narrative synthesis cannot keep pace with the volume and conceptual overlap of papers.

Method

LR‑Robot combines expert‑designed, multidimensional taxonomies with large language models (LLMs) that classify abstracts under carefully engineered prompts. Human‑in‑the‑loop evaluation iteratively refines taxonomies, prompts, and model choice. Retrieval‑augmented generation (RAG) stores classified records for downstream analysis such as temporal trend tracking and label‑enhanced citation networks.

Results

Applied to 12,666 option‑pricing articles, the framework defined a four‑dimensional taxonomy and benchmarked up to eleven mainstream LLMs. Expert‑constraint prompts raised F1 scores above 0.81 for binary model‑development classification and achieved high consistency across runs. Multi‑label dimensions (underlying asset, option type, model type) yielded sample F1 scores >0.82, with Gemini Flash 3.0 selected as the primary model. Classification enabled detailed temporal evolution plots and label‑enhanced citation analyses, revealing stable modeling share, rising exotic‑option research, and a surge of machine‑learning methods.

Implications

LR‑Robot demonstrates that a human‑in‑the‑loop LLM pipeline can accelerate labor‑intensive SLR stages while preserving interpretive accuracy. The approach is adaptable to other domains with dense terminology, and the enriched knowledge base supports richer bibliometric insights than traditional metadata‑only methods.