Standing benchmark
Fixed datasets, fixed metrics, and every method scored the same way.
Cross-modal integration comparison
Transfer accuracy is the headline number most integration papers report. The modality probe column is the diagnostic: a linear classifier trained to recover the source modality from the integrated representation. High transfer accuracy together with high probe accuracy means the modalities point in the same direction without sharing a representation.
Why both columns
A method can score well on mixing and transfer while a probe still recovers the source modality almost perfectly. Reporting transfer accuracy alone hides that. Running the probe across every method is what turns the observation into a property of the field rather than a quirk of one model.