PreScience : A Benchmark for Forecasting Scientific Contributions

Anirudh Ajith, Amanpreet Singh, Jay DeYoung, Nadav Kunievsky, Austin C. Kozlowski, Oyvind Tafjord, James Evans, Daniel S Weld, Tom Hope, Doug Downey · 2026

Read PreScience : A Benchmark for Forecasting Scientific Contributions by Anirudh Ajith, Amanpreet Singh, Jay DeYoung, Nadav Kunievsky, Austin C. Kozlowski, Oyvind Tafjord, James Evans, Daniel S Weld, Tom Hope, Doug Downey (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-20459.txt.

Description

Can AI trained on the science record foresee the gains that follow? Anirudh Ajith and nine co-authors offer PreScience, a dataset and benchmark for science forecasting. It builds from ninety-eight thousand recent AI research papers, plus linked papers on author records and links for five hundred two thousand papers. Records include titles, abstracts, named authors, key sources, topic labels, cite paths, and details frozen at time cutoffs.

Seven sample tasks cover work generation, partner forecast, prior work choice, cite-count forecast, future mix forecast, and two topic trend outlooks. Starting points range from simple rules to frontier models and agentic systems, with a new LACER metric judging made contributions more like humans do. Building task models into a twelve-month made-up corpus yields papers clearly less diverse and less novel than human research from the same period.

Catalog record

CREATORAnirudh Ajith, Amanpreet Singh, Jay DeYoung, Nadav Kunievsky, Austin C. Kozlowski, Oyvind Tafjord, James Evans, Daniel S Weld, Tom Hope, Doug Downey
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE191 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-20459

Browse the full library · About the Tower of Babel