DRIVE Improves Auto-Bidding in Real-Time Advertising with New Framework

Auto-bidding is a critical component in real-time advertising, requiring decisions that optimize long-term performance while adhering to budget and cost constraints. Existing methods, including offline reinforcement learning and Transformer-based sequence modeling, often struggle with sparse or long-tail traffic, leading to suboptimal averaged actions. To address these limitations, a new unified Transformer-based framework called DRIVE (Distributional and Retrieval-Augmented Bidding with Value Evaluation) has been introduced. DRIVE decouples the generation of candidate actions from the final decision-making process for offline auto-bidding. The framework integrates three key components: distributional action modeling, retrieval-augmented candidate generation from high-quality historical decisions, and value-based evaluation to select the most promising bid. Extensive experiments on various benchmarks demonstrate that DRIVE consistently improves bidding performance and generalizes well across multiple Transformer-based methods.