GUIDE: A Generative Framework for Safer Auto-Bidding in Digital Advertising
Researchers have introduced GUIDE (Generative Auto-Bidding with Unified Modeling and Exploration), a novel framework that combines directed exploration with a secure fallback option for automated bidding. Utilizing a Decision Transformer (DT), GUIDE concurrently models past bidding behaviors and transitions in environmental states, while a Q-value module facilitates exploration through regularization constraints. To ensure safety, an inverse dynamics model is employed. This method overcomes the shortcomings of previous approaches: rule-based systems were inflexible, Reinforcement Learning faced challenges with long-term dependencies, and current generative models lacked clear safety protocols, resulting in ineffective exploration and financial risks. The study is available on arXiv with the identifier 2605.19457.
Key facts
- GUIDE stands for Generative Auto-Bidding with Unified Modeling and Exploration.
- The framework uses a Decision Transformer to model bidding actions and state transitions.
- A Q-value module guides exploration through regularization constraints.
- An inverse dynamics model provides a safe fallback mechanism.
- The approach addresses inefficiencies in prior generative models for bidding.
- The paper is available on arXiv with ID 2605.19457.
- Automated bidding is central to modern digital advertising.
- Reinforcement Learning approaches modeled bidding as a Markov Decision Process.
Entities
Institutions
- arXiv