A new study from researchers at the University of Washington and the Allen Institute for AI (Ai2) has introduced OpenScholar, an artificial intelligence system that can synthesize scientific literature and cite sources with accuracy comparable to human experts. Developing tools to help researchers keep up with the millions of new scientific papers published every year has been a major challenge, because general-purpose AI models often hallucinate — inventing or misattributing citations rather than grounding responses in real research. OpenScholar was built specifically to overcome this problem.
To train and evaluate OpenScholar, the team assembled a corpus of about 45 million open-access scientific papers and used a technique called retrieval-augmented generation (RAG), which lets the model search, retrieve and integrate up-to-date research into its answers. The researchers also created a new evaluation benchmark — ScholarQABench — with thousands of queries and expert-written answers spanning fields like computer science, physics, biomedicine and neuroscience, in order to test how well AI systems actually understand and cite complex literature.
In side-by-side comparisons with state-of-the-art models such as OpenAI’s GPT-4o, OpenScholar showed substantially better performance. GPT-4o routinely fabricated 78–90 % of its citations in earlier tests, whereas OpenScholar’s citations were as reliable as those written by human researchers. In human evaluations, scientists preferred OpenScholar’s responses over expert-written answers about half the time, and combining OpenScholar’s citation methodology with larger models further improved preference rates substantially.
The team’s findings, published in the journal Nature, suggest that AI can now genuinely assist scientists in literature review and research synthesis without misleading citation errors, a longstanding limitation of earlier systems. Because the code, data and a live demo are open-source and freely available, other researchers can build on this work for broader scientific use and possibly extend it to support more comprehensive and multi-step research tasks in the future.
