SWAG: sliding-window-aware generative autobidding¶
Fidelity: core mechanism. Executes a proxy for the window-aware bidding objective and constraints. The public offline experiment is not a production ad system.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2607.25233 |
| Organization | Alibaba International Digital Commerce / Dalian University of Technology |
| First public release | 2026-07-28 (arXiv v1) |
| Authors' code | The paper did not supply official or author code when checked on 2026-08-09. |
| Adapter / method | swag-bid |
| Local reproduction code | src/auto_research/reproductions/swag_bid/ |
Original paper¶
Motivation and method¶
A masked future plan models a seven-day sliding-window objective across episodes. A per-step gate injects long-horizon guidance into the current bid decision.
flowchart LR
H["Campaign history"] --> M["Masked future plan"]
W["Seven-day window objective"] --> M
M --> G["State gate"]
G --> B["Bid or action score"]Key figure from the paper¶
Figure 2: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.
Core equations¶
Paper results¶
In a 21-day campaign-randomized AliExpress A/B test: cost +1.96%, GMV +3.42%, ROAS +5.65%, and goal attainment +2.02 percentage points.
Local reproduction¶
Local comparison: The baseline is a single-episode Decision Transformer proxy and the treatment is a sliding-window planner. On the same small candidate-ranking dataset, Hit@10 and NDCG@10 both change by +0.00%.
The run maps periodic MovieLens behavior to campaign windows and executes future masking, a seven-day MPC score, and a state gate.
See the stable result.
Reproduction boundaries¶
This is not a production Decision Transformer checkpoint. The public data contain no bid, budget, or GMV fields. These results are a structural ablation only.