Reproduce “DaV-Gen: End-to-End Generative Retrieval via Draft-and-Verify” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Meng Zhao, Chunmei Liu, Qinyong Wang
DaV-Gen introduces an end-to-end "Draft-and-Verify" framework for generative information retrieval, addressing objective inconsistency and inference latency in large-scale search and recommendation systems. It achieved a 2.5x inference speedup compared to traditional cascade systems and improved user engagement by +2.09% in online A/B tests while increasing user conversion by +0.47%.
Key insights
- Fusing sparse structural tokens with dense semantic vectors in a hybrid representation is crucial for achieving both efficiency in drafting and expressiveness for fine-grained item understanding.
- Progressive knowledge injection and a composite training loss (combining contrastive, generative, and pairwise objectives) effectively align the model's various sub-tasks and bridge the semantic gap for PLMs in domain-specific retrieval.
- Reframing generative retrieval from sequential auto-regressive decoding to a parallel 'Draft-and-Verify' mechanism is essential for achieving industrial-grade inference speeds and deterministic control over recommendation list length.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.