AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation¶

Fidelity: core mechanism. Paper online evidence, public-data experiments, and omissions are separate.

Paper information¶

Field	Details
Paper	arXiv 2603.19710
Organization	Taobao & Tmall Group / Alibaba
First public release	2026-03-20 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-08-24.
Adapter / method	`aigq`
Local reproduction code	`src/auto_research/reproductions/aigq/`

Original paper¶

Motivation and method¶

The Direct path keeps latency low and the Reasoning path handles complex intent. IL-GRPO optimizes whole query sets with online CTR reward, using hybrid offline and online serving.

flowchart LR
 A["Compress behavior context"] --> B["Direct and Reasoning"] --> C["IL-GRPO query list"]

Key figure from the paper¶

Figure 2: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equation¶

\[ J=\mathbb E[R(Q)]-\beta D_{KL}(\pi_\theta\|\pi_{ref}) \]

Paper results¶

The paper reports GMV +10.68% in a thirty-day online A/B test (Section 5.4). Its private protocol is not directly comparable with local MovieLens metrics.

Local reproduction¶

Local comparison: The baseline is the shared transition + content scorer. Adding only aigq on the same users, items, full catalog, and seed yields +8.09% relative NDCG@10.

MovieLens-100K, 220 users and 360 items, seed 42: NDCG@10 0.0540 → 0.0584 (+8.09%), Hit@10 0.1091 → 0.1136. Validation only selects mixture weights; test does not tune.

auto-research reproduce --paper aigq --dataset-dir data --seed 42

See the fixed metrics.

Reproduction boundaries¶

The local core mechanism uses fixed full-catalog candidates, a shared MovieLens-100K split, and a shared seed. Private features, production models, and online traffic are unavailable. The implementation has independent model state and scoring, retains negatives, and separates local changes from paper A/B gains.