Scientific agents support literature‑based tasks such as retrieval, question answering, evidence‑grounded generation, and claim assessment. Most existing systems are built around single tasks, causing the same papers to be repeatedly fetched, segmented, and interpreted, and making knowledge gained in one task hard to reuse in another.
ScholarStack introduces a layered research‑asset framework that compiles a collection of papers into reusable, versioned, provenance‑preserving assets at three complementary levels: source‑grounded paper‑level statements, domain‑level organization, and evidence‑grounded cross‑paper syntheses. A unified access interface returns task‑specific views at the evidence granularity each task requires, preserving study conditions, source traceability, and verification status.
We instantiated the framework across four task families covering ten settings, comparing agents that leverage the compiled assets with task‑specific baselines under matched base models. Quality gains concentrate on tasks that need cross‑paper evidence, such as multi‑paper QA and literature review generation, while query‑time token cost drops for every measured task because assets are compiled once and reused.
These results suggest that layered research assets can serve as shared infrastructure for scientific agents, shifting literature assistance from isolated document processing toward cumulative, evidence‑grounded workflows.
Review