Efficiently updating DrIG for dynamic candidate collections

Investigate how DrIG can efficiently handle candidate insertion, deletion, and identifier updates in dynamic candidate collections without requiring expensive retraining.

Background

DrIG constructs discrete residual-quantized identifiers for multimodal candidates and uses those identifiers during generative retrieval. In practical retrieval systems, however, candidate collections may change over time as items are inserted or deleted, or as existing item representations and identifiers require revision.

The paper leaves unresolved how the DrIG framework can maintain an up-to-date retrieval index under these operations without retraining the model or rebuilding identifiers at prohibitive cost. This problem is important for deploying generative multimodal retrieval in continuously changing collections.