跳到正文
研踪PaperTrail
← 研踪 PaperTrail

arXiv:2611.25738 · v1

CACO: Coverage-Aware Coordinated Optimization for Continuous Generative Retrieval

首次提交 2026-11-19 16:48
版本更新 2026-11-19 16:48
雷达观测 2026-11-20 19:58
arXiv ↗PDF ↗

摘要

Continuous generative retrievers powered by flow matching and diffusion sample multiple queries in item embedding space for parallel approximate nearest-neighbor (ANN) search under fixed retrieval budgets. However, optimizing distribution diversity or target-level interest separation ignores set-level retrieval geometry: sampled queries collapse onto overlapping ANN neighborhoods, while discrete top-K lookup and single-target supervision block end-to-end coverage optimization. We introduce CACO (Coverage-Aware Coordinated Optimization), a lightweight self-supervised post-training framework that coordinates queries from frozen flow-matching and diffusion retrievers toward complementary item coverage with negligible additional inference latency. CACO uses empirical retrieval frequencies across sampled reference queries as a relevance proxy, formulating a label-free coverage objective that improves recall across all 30 backbone-dataset settings in offline greedy selection experiments. To avoid the inference overhead of generating a large query pool, CACO employs a coordinated query adjustment network that combines inter-query bidirectional attention with gated relative directions, finally updating queries by bounded tangent-space projection and spherical retraction. During training, CACO applies tangential antithetic perturbations and measures each query's marginal gain via counterfactual single-query replacement, yielding an unbiased gradient estimator of smoothed set coverage across non-differentiable ANN search. On eight public benchmarks and an in-house 24-partition streaming industrial dataset, CACO improves 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.

作者备注:21 pages, 16 figures, 12 tables

AI 论文解读

基于摘要 · 非全文审阅

此版本尚未生成 AI 解读。解读由后台任务生成并按论文版本保存,登录后共享已有结果。

版本差异

目前只收录了一个版本。收录下一版后,可以对照标题、作者、分类和摘要。

观测到的版本

相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。

v1 · 2026-11-19 16:48 · 最新观测版本

CACO: Coverage-Aware Coordinated Optimization for Continuous Generative Retrieval

Continuous generative retrievers powered by flow matching and diffusion sample multiple queries in item embedding space for parallel approximate nearest-neighbor (ANN) search under fixed retrieval budgets. However, optimizing distribution diversity or target-level interest separation ignores set-level retrieval geometry: sampled queries collapse onto overlapping ANN neighborhoods, while discrete top-K lookup and single-target supervision block end-to-end coverage optimization. We introduce CACO (Coverage-Aware Coordinated Optimization), a lightweight self-supervised post-training framework that coordinates queries from frozen flow-matching and diffusion retrievers toward complementary item coverage with negligible additional inference latency. CACO uses empirical retrieval frequencies across sampled reference queries as a relevance proxy, formulating a label-free coverage objective that improves recall across all 30 backbone-dataset settings in offline greedy selection experiments. To avoid the inference overhead of generating a large query pool, CACO employs a coordinated query adjustment network that combines inter-query bidirectional attention with gated relative directions, finally updating queries by bounded tangent-space projection and spherical retraction. During training, CACO applies tangential antithetic perturbations and measures each query's marginal gain via counterfactual single-query replacement, yielding an unbiased gradient estimator of smoothed set coverage across non-differentiable ANN search. On eight public benchmarks and an in-house 24-partition streaming industrial dataset, CACO improves 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.

查看此版本解读 →查看此版本原文 ↗