Multimodal indexes can recreate source images

Research shows dense multi-vector visual document indexes can be inverted to reconstruct original pages. On the ViDoRe v3 benchmark, reconstruction recovers 47% of words and ranks the source document first in 98.4% of queries. Token-pooling and shuffling cut leakage to about 8% word recall, but shuffling is defeated by an order-restoring model.

• Index inversion recovers 47% of page words
• Source document ranks first in 98.4% of queries
• Token-pooling and shuffling cut recall to about 8%
• Order restoration lifts top-1 share from 3.8% to 93.5%