How does a new retrieval framework align user preferences with search results without slowing down?

What generative retrieval misses today

Generative retrieval maps a query straight to a semantic identifier for a candidate item. The identifier is built from item content alone. That creates a gap: the identifier stays static while the query intent shifts with each user. Current systems also skip user behavior sequences. They rely on beam search with an autoregressive decoder, which adds latency at every step.

Submitted 31 Aug 2026. Findings paper at EMNLP 2026. Twenty-two pages, ten figures, seven tables. Seven authors: Gaoming Zhang, Angqing Jiang, Jianchun Song, Kena Qi, Dayao Chen, Wei Lin, Defu Lian.

How CHAP aligns hierarchy and behavior

CHAP stands for Cross-component Hierarchical semantic Alignment for Personalized generative retrieval. First, a Hierarchical Semantic Alignment module matches the query latent space to the item quantization path. This synchronizes semantics at multiple granularities. Second, the framework models user behavior by combining discrete semantic identifiers for structural guidance with continuous representations for fine-grained refinement. The two streams work together instead of replacing each other.

The single-pass decoder trick

A Residual Cascading Generation mechanism confines the multi-step Transformer decoder to one forward pass. The design keeps throughput high while limiting information loss. The authors say this removes the bottleneck that beam search creates in standard generative retrieval.

What the experiments show

Experiments cover three public datasets, one proprietary industrial dataset, and online A/B tests. The paper reports superiority across these settings. Code is publicly available. The document does not specify how many items or users the industrial dataset contains. It also does not report latency numbers in absolute milliseconds.