Iterative Scarcity-Guided Exploration: Bootstrapping Generative Auto-bidding from Narrow Support

Auto-bidding is a core algorithmic component in large-scale auctions. In industrial cold-start settings, iterative training is commonly used to compensate for low-quality, narrowly supported historical data. Representative AI-Generated Bidding (AIGB) methods based on conditional diffusion planners have shown strong empirical performance. However, diffusion planners fail to sustain improvement in iterative training, becoming trapped in a suboptimal performance bottleneck. In this paper, we theoretically attribute this stagnation to the collapse of the effective signal-to-noise ratio (SNR): the sparse radial return signal is overwhelmed by a curvature-induced penalty, preventing extrapolation beyond the narrow support. To break this limit, we propose \textbf{Iterative Scarcity-Guided Exploration (ISGE)}, which alternates between a \emph{Judger} that detects high-scarcity trajectories and an \emph{Explorer} that performs guided generation toward high-scarcity and high-return regions. We theoretically show that ISGE elevates the effective SNR above its critical threshold and reshapes the optimization landscape, enabling the policy to extrapolate beyond the support. Extensive experiments on the industrial AuctionNet benchmark demonstrate that ISGE effectively bootstraps from narrow support and surpasses the performance of full-dataset baselines within three iterations.