A new preprint on arXiv examines whether calibration can make up for the absence of negative passages when training answerability readouts—models that judge whether a passage actually supplies an answer. The abstract frames the question as unresolved: it is not clear if calibration compensates for removing such negatives during training.
The paper presents a controlled study of how training-negative passages affect the readout. While the abstract is truncated, the title states the conclusion: training negatives still matter after calibrating answerability readouts. This suggests that calibration alone does not fully substitute for explicit negative examples in the training data.
As a single source with a partial abstract, the details of the experimental setup and effect sizes are not available here. The finding, if confirmed, would be relevant for practitioners who hope to simplify training by dropping negative passages and relying on post-hoc calibration instead.