- A study published in Nature on February 4, 2026, reveals that the OpenScholar chatbot can outperform PhDs and postdocs in writing scientific literature reviews.
- OpenScholar was designed by U.S. scholars to overcome the “hallucinations” commonly found in LLMs like ChatGPT or Llama when citing research papers.
- Subject matter experts in computer science, physics, neuroscience, and biomedicine compared reviews written by OpenScholar, ScholarQABench, and doctoral students.
- Results showed that OpenScholar was preferred in 51% of cases, while ScholarQABench reached 70%, significantly higher than reviews written by PhDs.
- The main advantage comes from coverage and depth of information, with reviews averaging 1,447 words (OpenScholar) or 706 words, compared to 424 words by humans.
- Summaries generated by ChatGPT were preferred in only 31% of cases due to a lack of comprehensive content.
- The study pointed out that popular LLMs generate fake citations in 78–90% of cases, and even 78–98% of document titles are fabricated, which is particularly severe in biomedicine.
- Conversely, OpenScholar recorded no hallucinations in computer science and biomedicine reviews.
- The 8B OpenScholar model was trained on 45 million scientific papers, creating a self-improving feedback loop for citation accuracy.
- Since its demo launch, OpenScholar has reached over 30,000 users with nearly 90,000 queries; the cost per review ranges from only $0.01 to $0.05.
📌 Conclusion: A study in Nature (Feb 4, 2026) shows that the OpenScholar chatbot, with only 8 billion parameters and trained on 45 million papers, can outperform PhDs and postdocs in scientific literature reviews. Its edge lies in information coverage and depth, averaging 1,447 or 706 words versus 424 words from humans. While popular LLMs fabricate citations in 78–90% of cases, OpenScholar recorded zero hallucinations in computer science and biomedicine.
