AI-Generated Bidding (AIGB): Using Generative Models for Automated Advertising Bidding

Recent advances in generative large models such as ChatGPT have sparked a new wave of interest in the advertising industry. These models are expected to reshape user‑product interaction and consequently disrupt traditional advertising and automated bidding systems.

The evolution of automated bidding at Alibaba can be divided into three stages: (1) budget‑consumption control using classic PID controllers, (2) RL‑based bidding that treats bidding as a sequential decision problem, and (3) Sustainable Online RL (SORL) which learns directly in the online environment to avoid simulation‑reality gaps.

Building on this background, the Alibaba‑Mama team proposes AIGB (AI Generated Bidding), a conditional generative‑model approach that replaces the reinforcement‑learning perspective. Instead of learning a policy as a black‑box action, AIGB models the joint distribution of bids, objectives, and constraints and generates a conditional distribution of bidding strategies given the desired goals and limits.

Training uses historical sub‑optimal bidding trajectories as samples and fits the joint distribution via maximum‑likelihood estimation. At inference time, the model samples bid strategies that satisfy the specified constraints (e.g., CPC, ROI, smoothness) while optimizing the chosen objective (e.g., click‑through rate, conversion volume).

The architecture consists of a conditional diffusion or transformer backbone that processes the current trajectory and conditioning variables, then sequentially generates future bid actions. This design offers better interpretability, longer‑horizon planning, and smoother feedback compared with traditional RL methods.

Advantages of the generative approach include: (1) avoidance of distribution shift and policy degradation during training, (2) ability to generate diverse strategies for multiple constraint combinations, (3) improved explainability and easier integration of expert knowledge, and (4) higher efficiency in model iteration.

In summary, AIGB demonstrates that generative AI can fundamentally restructure automated bidding pipelines, providing a scalable solution for large‑scale advertising platforms. The authors anticipate further improvements as more massive ad‑delivery data become available for training even larger decision‑making models.