An open-source tool called OpenScholar, published in Nature, promises to make literature reviews cheaper and more reliable by grounding answers in a 45-million-article open-access database and linking claims directly to sources. The model matched human experts on citation accuracy and outperformed several large commercial LLMs, while costing a fraction of proprietary “deep research” add-ons. Researchers can run it locally or adapt its method to other models, though coverage limits and imperfect retrieval remain constraints. The release arrives amid mounting concerns about fabricated citations in AI-assisted research, including flawed references found in submissions to major machine-learning conferences.
Related articles:
— Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
— Galactica: A Large Language Model for Science
— S2ORC: The Semantic Scholar Open Research Corpus
— CORD-19: The COVID-19 Open Research Dataset
— Toolformer: Language Models Can Teach Themselves to Use Tools
