In a significant breakthrough for the scientific community, researchers from the University of Washington and The Allen Institute for AI (Ai2) have developed an artificial intelligence model, OpenScholar, that matches human accuracy in citing scientific research. This development is crucial as scientists grapple with the challenge of staying updated amidst the deluge of millions of scientific papers published annually.
Artificial intelligence systems have long promised to assist in synthesizing vast amounts of information. However, they often falter by generating inaccurate or fabricated data, commonly referred to as “hallucinations.” A study of OpenAI’s GPT-4o model revealed that it fabricated 78-90% of its research citations. Furthermore, general-purpose AI models like ChatGPT struggle to access papers published after their training data was collected.
OpenScholar: A Solution to AI’s Citation Challenges
In response to these limitations, the UW and Ai2 research team developed OpenScholar, an open-source AI model tailored to synthesize current scientific research. They also introduced the first large, multi-domain benchmark to evaluate models’ ability to synthesize and cite scientific research accurately. In rigorous tests, OpenScholar’s citation accuracy matched that of human experts, with 16 scientists preferring its responses over those written by subject experts 51% of the time.
The findings were published on February 4 in the journal Nature. The project’s code, data, and a demo are publicly available, underscoring the team’s commitment to transparency and collaboration.
“After we started this work, we put the demo online and quickly, we got a lot of queries, far more than we’d expected,” said senior author Hannaneh Hajishirzi, a UW associate professor and senior director at Ai2. “When we started looking through the responses, we realized our colleagues and other scientists were actively using OpenScholar. It really speaks to the need for this sort of open-source, transparent system that can synthesize research.”
Groundbreaking Techniques and Testing
The researchers trained OpenScholar using a dataset of 45 million scientific papers, ensuring its answers were grounded in established research. They employed a technique known as “retrieval-augmented generation,” allowing the model to search for new sources, incorporate them, and cite them post-training.
Lead author Akari Asai, a research scientist at Ai2, explained their approach: “Early on, we experimented with using an AI model with Google’s search data, but we found it wasn’t very good on its own. It might cite some research papers that weren’t the most relevant, or cite just one paper, or pull from a blog post randomly. We realized we needed to ground this in scientific papers.”
To evaluate their system, the team developed ScholarQABench, a benchmark for testing systems on scientific search. This involved gathering 3,000 queries and 250 longform answers written by experts across various fields, including computer science, physics, biomedicine, and neuroscience.
“AI is getting better and better at real-world tasks,” Hajishirzi noted. “But the big question ultimately is whether we can trust that its answers are correct.”
Performance and Future Prospects
OpenScholar was tested against other leading AI models, such as OpenAI’s GPT-4o and two models from Meta. ScholarQABench automatically evaluated the AI models’ answers based on metrics like accuracy, writing quality, and relevance. OpenScholar outperformed all the systems it was tested against.
When scientists compared answers from the models with human-written responses, they preferred OpenScholar’s answers 51% of the time. Notably, when OpenScholar’s citation methods were combined with GPT-4o, a larger model, scientists favored the AI-written answers 70% of the time, while GPT-4o alone was preferred only 32% of the time.
“Scientists see so many papers coming out every day that it’s impossible to keep up,” Asai said. “But the existing AI systems weren’t designed for scientists’ specific needs. We’ve already seen a lot of scientists using OpenScholar, and because it’s open-source, others are building on this research and already improving on our results.”
The team is now working on a follow-up model, DR Tulu, which builds on OpenScholar’s findings and performs multi-step search and information gathering to produce more comprehensive responses.
Collaborative Efforts and Broader Implications
The development of OpenScholar involved a collaborative effort from a diverse group of researchers, including Jacqueline He, Rulin Shao, Weijia Shi, Dan Weld, Varsha Kishore, Luke Zettlemoyer, Pang Wei Koh, and others from institutions such as the University of Illinois Urbana-Champaign, University of North Carolina, Stanford University, and Carnegie Mellon University.
This project highlights the potential of AI to revolutionize how scientists access and synthesize information, addressing a critical need in the research community. As AI models continue to evolve, their ability to provide accurate and reliable information will be pivotal in advancing scientific discovery.
Looking ahead, the success of OpenScholar sets a precedent for future AI developments in research and academia, emphasizing the importance of open-source collaboration and continuous improvement.
