Xin Xin

Generative recommendation's cold-start gains are often illusory, inflated by inconsistent evaluation and confounding design choices like model scale and identifier design.

Forget retraining: DOME efficiently adapts generative retrieval models to new documents by directly editing the model's docID mapping, achieving comparable performance with 40% less training.

Achieve 97x speedup and 208x energy efficiency for Multi-Scale Deformable Attention by co-designing a near-memory processing architecture that overcomes the limitations of existing NMP solutions.