How Bid2X Revolutionizes Online Ad Bidding with a Universal Foundation Model

Abstract

Automatic bidding services are essential for advertisers, yet existing methods are often tailored to a single scenario and fail to generalize. Bid2X proposes a universal bidding‑environment foundation model (BFM) that learns the relationship between a given bid and its possible outcomes (e.g., budget consumption, GMV, PV) across diverse scenarios.

1. Introduction

Current auto‑bidding algorithms implicitly model the bidding environment but struggle with comprehensive understanding and cross‑scenario generalization. Challenges include heterogeneous data formats, complex dynamic dependencies, and a zero‑inflated distribution of bidding outcomes.

2. Preliminaries

We define automatic bidding and bidding‑environment modeling, introduce the concept of a foundation model for advertising, and formalize the bidding‑environment prediction problem as a self‑supervised next‑token prediction task.

3. Method

Bid2X consists of three main components:

Unified Data Embedding : Heterogeneous historical and today’s bidding data are encoded into a unified sequence representation. Historical data are embedded per variable, while today’s data are treated as tokens per time step, with position encodings added.

Variable and Time Attention Transformers : Variable attention captures pairwise relationships among variables, and causal time attention models temporal dynamics. Both attention outputs are fused by a variable‑aware fusion module.

Zero‑Inflation Projection & Auxiliary Tasks : A binary classifier predicts the probability of a non‑zero target, enabling a zero‑inflated projection. An auxiliary cumulative‑prediction task provides a global view of the bidding environment.

The overall loss combines zero‑inflation cross‑entropy and cumulative mean‑square error:

Loss = CE(zero prob) + MSE(cumulative prediction)

4. Experiments

Datasets & Baselines : Eight real‑world advertising datasets from the Taobao platform (≈100 M trajectories, 3 M records) were used. Baselines include Informer(fm) and other state‑of‑the‑art models.

Metrics : Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were reported.

Results : Bid2X consistently outperformed baselines across all datasets. Ablation studies showed that removing variable attention (w/o va) or time attention (w/o ta) caused the largest performance drops, confirming their necessity.

Scalability : Performance improved predictably with increasing model size N and data size D, demonstrating good sample efficiency.

Online A/B Test : Deployed on Taobao’s live ad system, Bid2X increased GMV by 4.65 % and ROI by 2.44 % compared with a reinforcement‑learning baseline.

5. Conclusion

Bid2X establishes a foundation‑model‑based approach for bidding‑environment modeling, achieving strong cross‑scenario generalization, handling zero‑inflated data, and delivering measurable business impact. The work has been accepted to the KDD'25 ADS Track.

Figures

References

D. Moriwaki et al., “A real‑world implementation of unbiased lift‑based bidding system,” IEEE Big Data, 2021.

H. Cai et al., “Real‑time bidding by reinforcement learning in display advertising,” WSDM, 2017.

Z. Mou et al., “Sustainable online reinforcement learning for auto‑bidding,” NeurIPS, 2022.

X. Li et al., “Arbitrary distribution modeling with censorship in real‑time bidding advertising,” KDD, 2022.

R. Bommasani et al., “On the opportunities and risks of foundation models,” arXiv, 2021.