Scientific agents must bridge research problems, methods, and findings that are scattered across narrative articles. To address this, we introduce SPRA (Structured Paper Representation for Agents), a paper‑level model that treats research objects as persistent entities linked by typed relations. SPRA’s empirical‑AI schema organizes Problems, Contributions, Components, ExperimentSetups, Measures, and Findings into queryable, traversable, and reusable units, while selected citation relations capture dependencies, component usage, and comparisons with prior work, all anchored to source locations for verification.
We instantiated SPRA on more than one million AI papers, yielding roughly 19.95 million research units and 24.07 million internal relations. Accessing SPRA through Schema‑RAG, with task evidence fixed before inference, improves source‑conditioned scientific QA by 0.381/5 over passage retrieval and stays within 0.078/5 of Full Compact on average. In directed cross‑paper reasoning, qualification accuracy rises by 0.184/5 over passage retrieval; Top‑5 synthesis shows mixed quality effects. Across all paired RQ2 tasks (including Top‑5), input + output tokens drop 28.5% relative to passage retrieval and 95.3% relative to Full Compact. In idea generation, the overall score is only 0.067 points lower than Full Compact while using 93.7% fewer tokens. These results demonstrate that SPRA is a usable representation of scientific content and a foundation for workflows that connect, inspect, and reuse paper‑level research objects. The SPRA resource is available via the SciVerse scholarly‑data platform.
Review: By converting scattered scholarly information into structured entities and explicit relations, SPRA creates a machine‑readable graph that dramatically reduces retrieval and reasoning token costs while preserving or improving QA and reasoning performance, paving the way for agent‑centric research assistants.