Scientific discovery is an iterative and knowledge-intensive process that heavily relies on the effective retrieval of prior literature. However, many existing Scientific Document Retrieval (SDR) benchmarks do not adequately capture the complexity of ...
Scientific discovery is an iterative and knowledge-intensive process that heavily relies on the effective retrieval of prior literature. However, many existing Scientific Document Retrieval (SDR) benchmarks do not adequately capture the complexity of real-world research, suffering from limited scalability due to static corpus snapshots and a lack of explicit modeling of citation intents, that is, the reasons why researchers cite specific papers. To address these limitations, this thesis introduces the IntentSDR Framework, an LLM-based framework for scalable, citation-intent-aware SDR benchmarking grounded in citation relations, and instantiates it as the IntentSDR-Bench.
First, we propose a scalable seed-expansion strategy that traverses the citation network to construct stratified corpora ranging from thousands to hundreds of thousands of documents, allowing us to evaluate retrieval systems under continuously growing literature. Second, we define a hierarchical taxonomy of 11 citation intents, explicitly incorporating result-centric categories such as negative or null results to capture functional relationships that are often overlooked in prior work. Third, we implement a cost-effective LLM-driven pipeline that extracts these intents from citation contexts and formulates structured queries using specialized templates, thereby ensuring high domain relevance while reducing hallucinations.
Extensive experiments with various retrieval models (sparse, dense, and LLM-based) show that IntentSDR-Bench maintains consistent performance characteristics across different corpus scales, demonstrating system-level scalability. Furthermore, an "LLM-as-a-judge" evaluation confirms the high quality of the generated queries in terms of evidence support, self-containment, and intent alignment, while achieving significant cost efficiency without compromising query quality. Taken together, the IntentSDR Framework provides the research community with a sustainable tool for evaluating next-generation SDR systems under rich, intent-driven scientific information needs.