CUSP is a research dataset for evaluating how well AI systems handle temporally grounded questions about scientific papers under a historical cutoff. Given a model's training cutoff date, CUSP asks: can the model anticipate real scientific discoveries that happened after that date?
Authors: Sean Wu*, Pan Lu*, Yupeng Chen, Jonathan Bragg, Yutaro Yamada, David Clifton, Philip Torr†, James Zou†, Junchi Yu†… See the full description on the dataset page:
https://huggingface.co/datasets/SeanWu25/CUSP.
