Multi-query retrieval stops duplicating itself: CACO coordinates sampled queries for complementary coverage, lifting Recall@100 by up to 23.7%
Synopsis
Continuous generative retrievers sample several query vectors for one user and run parallel approximate nearest-neighbor search, but the queries often land in overlapping neighborhoods and waste a fixed retrieval budget. CACO is a lightweight self-supervised post-training framework that freezes the flow-matching or diffusion generator, uses retrieval frequencies of reference queries as a relevance proxy to build a label-free set-coverage objective, and trains a coordination network that adjusts the queries jointly. Across eight public benchmarks and an in-house 24-partition streaming industrial dataset, CACO raises Recall@100 by up to 23.7% and Recall@400 by up to 23.9% over the strongest baseline, cuts inter-query overlap by 53.3%–64.9%, and adds only 0.5%–3.0% inference latency.
Interpretation
The value of multi-query continuous generative retrieval depends on how complementary the query set is, not only on how well each single query matches a target; CACO moves the optimization target from per-query relevance to the coverage of the merged set. Earlier continuous generative retrievers either sampled queries independently or learned separated interest representations from observed positive interactions, without explicitly accounting for set-level retrieval geometry, so even a diverse query distribution could yield heavily overlapping nearest-neighbor results. On eight public benchmarks and an in-house 24-partition streaming industrial dataset, CACO raises Recall@100 by up to 23.7% and Recall@400 by up to 23.9% over the strongest baseline, and cuts mean pairwise Jaccard overlap between queries by 53.3%–64.9%.
CACO does not retrain the generator; it adds a coordination network on top of a frozen flow-matching or diffusion backbone so each query is generated with awareness of what the other queries may retrieve. The coordination network applies cross-attention to user history and bidirectional self-attention across queries, uses a signed relative-directions gate to decide whether each query moves toward or away from the others, and keeps unit norm through bounded tangent-space projection and spherical retraction. Training signal comes from tangential antithetic perturbations and counterfactual single-query replacement: holding the other queries fixed, only one query's positive and negative perturbed versions are swapped, and the difference in their coverage gain estimates the gradient, bypassing non-differentiable approximate nearest-neighbor search; on Beauty, Sports, Toys and the industrial data, end-to-end inference time rises by only 0.51%–2.97%.
The coverage objective needs no ground-truth target labels; it weights items by how often reference queries sampled from the frozen generator retrieve them, and an item already covered yields no further gain when retrieved again. Offline greedy selection experiments first test whether this objective can pick higher-recall query subsets before any new model is trained. Across all 30 backbone–dataset settings formed by five retrieval backbones and six datasets, greedy optimization of the frequency-weighted coverage objective beats random selection while lowering inter-query overlap.
Perspective
The demonstrated scope is eight public benchmarks, an in-house 24-partition streaming industrial dataset and six frozen flow-matching and diffusion backbones, with the retrieval budget fixed at 20 queries and 20 or 400 neighbors per query; the industrial dataset is kept proprietary and platform-specific details are omitted for data security and commercial confidentiality.
On ML-25M at Recall@100, LLaDARec reaches 0.327 versus CACO's 0.305, and CACO only slightly exceeds it at Recall@400 (0.518 vs. 0.517); reference-pool refresh improves consistently under the retrieval setting used for training, but when transferred to a different query count and retrieval depth, mean Recall@100 improves on Beauty, Toys and KuaiRec while decreasing on Sports and ML-25M; further refresh rounds show diminishing returns and plateau.
