Skip to main content
Back to timeline
arXivSource publication:

Generative search stops treating the query as just another input: IntentGR lifts Recall@10 by 8.11% on AOL4PS and 11.96% on AliExpress search logs

Synopsis

End-to-end generative retrieval for personalized search has largely treated the query as one input feature, leaving identifier learning and sequence generation without explicit intent guidance, so generated semantic identifiers drift toward high-frequency token patterns. IntentGR adds query-level behavioral evidence to the input and pairs it with alignment and modulation mechanisms at the optimization, training-distribution and decoding stages. On the public AOL4PS benchmark and a large-scale AliExpress e-commerce dataset, Recall@10 improves by 8.11% and 11.96% over the strongest baseline, with especially large gains on long-tail queries.

Source-provided article image: IntentGR: Enhancing Intent for End-to-End Personalized Generative Retrieval in Search
arXiv

Interpretation

End-to-end generative retrieval for personalized search now carries query intent through several stages of the retrieval pipeline instead of only concatenating the query into the encoder input. Earlier end-to-end methods from recommendation focus on user-item interactions and leave query semantics underexplored, while search-oriented methods encode the query into the identifier space but treat it only as an input signal for aligning and balancing ID assignments and disregard user behavioral histories entirely. On AOL4PS, the controlled baseline ETEGRec-Q, which only adds the query to the input, raises Recall@10 from 47.03% for ETEGRec to 69.88%, while IntentGR reaches 77.99% on the same benchmark, an 8.11% gain over that strongest baseline.

The approach strengthens query semantics at four complementary points: input, optimization, training distribution and decoding. Query-Aware Semantic Behavior Representation (QBR) adds a sequence of earlier clicks by other users under the same query as query-level behavioral evidence alongside user history in the encoder context; Query-Semantic Alignment (QSA) introduces a bidirectional contrastive objective between query representations and reconstructed target-item representations into the joint optimization of tokenizer and retriever; Long-Tail Query Enrichment (LQE) raises the occurrence and semantic diversity of low-frequency queries through repeated sampling and query rewriting; Query-conditioned Attention Modulation (QAM) injects query semantics into the context the decoder attends to, so every position carries query information. On AE, a large-scale e-commerce dataset built from one week of AliExpress click logs, IntentGR reaches 34.36% Recall@10 against 22.40% for ETEGRec-Q, a relative gain of 11.96%.

The gains come mainly from turning query-level behavioral evidence into a generalizable generation target rather than returning historical clicks directly. Used alone as a non-parametric retrieval baseline that returns items from the query's historical click sequence, QBR achieves 69.36% Recall@10 on AOL4PS, comparable to the 69.88% of the generative baseline ETEGRec-Q. Full IntentGR reaches 77.99% on the same benchmark, 8.63% above QBR standalone; on test samples where the target item does appear in the query's behavior history, ETEGRec-Q achieves 89.11% Recall@10 and IntentGR 99.72%.

Perspective

The demonstrated scope is personalized search: the public AOL4PS benchmark and AE, an industrial e-commerce dataset drawn from one week of AliExpress click logs and evaluated on the last hour of the final day; long-tail query enrichment is applied only to training data, leaving validation and test sets unchanged.

On samples where the target item does not appear in the query's behavior history, IntentGR's gain over ETEGRec-Q is smaller, 2.44% on AOL4PS and 1.22% on AE, since both models then rely on query text alone. Sensitivity to the QSA contrastive coefficient and the QAM fusion coefficient is analyzed on AOL4PS, and performance drops at boundary settings.