PaperQA2

PaperQA2 is an AI agent developed by FutureHouse that achieves superhuman performance on scientific literature search tasks. Designed for researchers, it is optimized for retrieving and summarizing information from the scientific literature, outperforming PhD and postdoc-level biology researchers in accuracy as measured by the LitQA2 benchmark — part of the LAB-Bench evaluation set. PaperQA2 represents a new paradigm for interacting with and extracting knowledge from scientific publications at a scale previously unavailable to human researchers.

Built as an open-source project with an accompanying research paper, PaperQA2 powers several downstream agents, including WikiCrow — which generates Wikipedia-style scientific summaries more accurate on average than human-curated Wikipedia articles — and ContraCrow, which identifies contradictions across the published literature to surface potential new discoveries.

Core Capabilities

• Retrieves and summarizes information from the scientific literature with superhuman accuracy
• Accesses a suite of tools to find relevant papers, extract useful information, and explore citation graphs
• Formulates well-grounded answers by synthesizing content across multiple sources
• Enables literature analyses at a scale not feasible for human researchers

WikiCrow: Wikipedia-Style Scientific Summaries

• An agent built on top of PaperQA2 that generates Wikipedia-style summaries of scientific topics
• Produces summaries judged by blinded PhD and postdoc-level biology researchers to be more accurate on average than existing human-curated Wikipedia articles
• Previously used with an earlier version of PaperQA to generate articles for all 20,000 genes in the human genome by combining information from 1 million distinct scientific papers
• With PaperQA2's improved accuracy, FutureHouse plans to regenerate Wikipedia articles for all 20,000 human genes and release them at wikicrow.ai, with a preview of 240 articles already available

ContraCrow: Contradiction Detection and Hypothesis Generation

• An agent built on top of PaperQA2 that evaluates every claim in a scientific paper to identify contradicting statements from elsewhere in the literature
• Grades identified contradictions on a Likert scale to filter out trivial discrepancies
• Finds an average of 2.34 contradicted statements per paper in a random subset of biology papers
• Designed to help surface new hypotheses and propose pivotal experiments by exploring meaningful disagreements between published findings

PaperQA2 is available as open-source code on GitHub, with a corresponding research paper accessible via paper.wikicrow.ai. FutureHouse, a registered 501(c)(3) nonprofit, continues to develop PaperQA2 and its associated agents as part of a broader mission to transform how scientific knowledge is discovered, synthesized, and applied.