Regarding the prevalent issue of "fabricating papers" in AI models within the academic field, a research team from the University of Washington and the Allen Institute for AI (AI2) has introduced a breakthrough solution—officially releasing the open-source AI model OpenScholar. This model excels in integrating cutting-edge research, significantly improving citation accuracy, and its generated content is even preferred by more than half of human experts.

For a long time, even top models like GPT-4o have faced a "hallucination" rate of 78% to 90% when handling academic citations. To tackle this pain point, OpenScholar takes an alternative approach, building a vast retrieval library containing 45 million academic papers. Using advanced Retrieval-Augmented Generation (RAG) technology, the model can access the latest published literature in real-time and provide responses in standard citation formats, completely eliminating the awkwardness of "seriously making things up."

In rigorous benchmark tests using ScholarQABench and expert double-blind reviews, OpenScholar's performance was impressive. Test results showed that in 51% of test cases, scientists preferred the responses generated by OpenScholar over those written by human experts. If combined with GPT-4o's citation mechanism, the preference rate of experts soared to 70% straight away.

Currently, the code, dataset, and demonstration version of OpenScholar are all available to the public. Not only does it provide a powerful tool for researchers, but it also sets a new benchmark for building a transparent and reliable academic AI ecosystem. The team stated that they will continue to iterate, launching a new model that supports multi-step retrieval and information aggregation, further empowering scientific research.