DaV-Gen's Draft-and-Verify mechanism unifies search and recommendation pipelines
End-to-end generative retrieval that combines speed of sparse drafting with LLM precision.
DaV-Gen tackles a fundamental flaw in modern industrial information retrieval systems: the multi-stage cascade architecture (coarse retrieval then fine ranking) suffers from optimization inconsistencies between stages, causing early errors to amplify and degrade final results. Existing end-to-end generative models can unify this pipeline but are bottlenecked by the auto-regressive nature of decoder-only architectures, making online serving impractically slow.
Instead, DaV-Gen adopts a 'Draft-and-Verify' mechanism, inspired by speculative decoding. The single model is trained with a composite loss function: a contrastive loss structures the embedding space for efficient vector-based drafting, and a fusion loss combines generative likelihood with vector similarity for precise verification. At inference, it first drafts a candidate set nearly instantly via vector similarity, then verifies those candidates using the richer fused score. This achieves the speed of traditional sparse indexing and the accuracy of generative models, all within one end-to-end model. Accepted at IJCAI'26, the paper demonstrates a new paradigm for both search and recommendation.
- Replaces multi-stage cascade architectures with a unified generative model, eliminating error propagation across stages.
- Combines contrastive learning (for fast vector drafting) and fusion loss (generative + similarity scoring) in a single training objective.
- Achieves both the speed of sparse retrieval and the precision of advanced generative models via two-phase inference (draft then verify).
Why It Matters
DaV-Gen could fundamentally speed up and improve accuracy in search engines, recommendation systems, and any large-scale retrieval pipeline.