Skip to content

REFUTE v3 preprint package

Latest

Choose a tag to compare

@connerlambden connerlambden released this 24 Jul 19:56

Interim citable package while the Zenodo DOI is being finalized.

REFUTE asks whether models can read science honestly. Public writeup: https://bgpt.pro/refute

Current board: 320 questions, 15 complete Truth Scores, leader 74.5 (Grok-4.2). Main finding: critique skill and calibration dissociate.

Assets:

  • preprint PDF
  • methods notes
  • leaderboard JSON

Dataset: https://huggingface.co/datasets/BGPT-OFFICIAL/refute