Researchers introduced EDB-Bio, a benchmark built from published studies in three biological fields, to test whether language models can judge evidence for scientific hypotheses. In the researchers’ tests, untrained models were systematically underconfident and scored about as well as a simple base-rate forecast; training them to give better-calibrated probabilities improved their results.
The researchers say their Ibis0 model selected experiments 2.5 times more efficiently than the best frontier model in a sequential experiment-selection task. They released the code, data and trained models.
Why it matters
For scientists using AI to help choose experiments, the results suggest that probability-calibration training may improve experiment selection, though the reported advantage is limited to the study’s task and has not been independently validated here.
