Good Perspective
A text model training objective in which a text encoder learns alignment with an image space by distinguishing among different descriptions of the same image (positive sample pairs). Compared with next-token prediction, it achieves cross-modal alignment more efficiently.
Timeline (last 90 days)
Averaged over 3 random seeds at the same learning rate, good views (two different descriptions of the same photo) reached 0.320, bad views (a single description with some words masked) reached 0.192, and next-word prediction reached 0.148.
The authors performed a small learning-rate search for each method, taking the best rate with single-seed runs, yielding 0.328 / 0.196 / 0.168, with the ranking unchanged, showing that the lead is not an accidental product of learning-rate tuning.
The retrieval metric R@1 (with a fitted rotation) shows a narrower gap: good views 0.084, bad views 0.063, next-word prediction 0.057.
All Facts (3)
Averaged over 3 random seeds at the same learning rate, good views (two different descriptions of the same photo) reached 0.320, bad views (a single description with some words masked) reached 0.192, and next-word prediction reached 0.148.
50%UnverifiedThe authors performed a small learning-rate search for each method, taking the best rate with single-seed runs, yielding 0.328 / 0.196 / 0.168, with the ranking unchanged, showing that the lead is not an accidental product of learning-rate tuning.
50%UnverifiedThe retrieval metric R@1 (with a fitted rotation) shows a narrower gap: good views 0.084, bad views 0.063, next-word prediction 0.057.
50%