We're excited to share that our paper on OpenScholar has been accepted to Nature.
Scientists need to stay current, but the pace of publishing is relentless, and today's general-purpose AI systems still struggle with a basic requirement for research: reliable grounding. They can write plausible summaries, but when you ask for supporting evidence, they often cite irrelevant work or invent sources outright.
OpenScholar is an open-source model designed specifically for synthesizing scientific literature with verifiable citations, so scientists can move faster without sacrificing trust. It was built by researchers at Ai2 and the University of Washington, with a focus on transparency and reproducibility.
OpenScholar pairs a model trained for scientific synthesis with retrieval-augmented generation (RAG). This allows it to search a large scientific corpus, incorporate relevant papers (including newer ones), and cite sources for the claims it makes. To ground answers in the literature, we constructed a corpus of 45 million open-access scientific papers and developed a full-text snippet index for OpenScholar to retrieve from, which we later made available through the Semantic Scholar API.
We also built ScholarQABench, the first large, multi-domain benchmark for evaluating systems on scientific synthesis and citation quality. ScholarQA-CS, the computer science portion of ScholarQABench, later evolved into ScholarQA-CS2, the long-form scientific QA benchmark now included in AstaBench.
OpenScholar shows that a careful approach to retrieval, ranking, and citation handling can significantly improve how useful and trustworthy an answer is in scientific settings.
From the start, our goal was to build a system that scientists can inspect, validate, and extend. OpenScholar's model checkpoints, retrieval index, and data are all publicly available and free to use, and a demo is available here.
OpenScholar was a step toward automated research assistants that show their work, not just sound convincing. It laid the groundwork for ScholarQA, which evolved into the report generation capability now available in Asta. We're continuing to build on these findings with Deep Research Tulu (DR Tulu), which expands the approach with multi-step search and information gathering for more comprehensive, long-form research reports.