Generative Response Model for Auto-Bidding Systems

A recent research article introduces the Generative Response Model (GRM) designed for auto-bidding systems, with the goal of enhancing advertiser value while adhering to budget limits and ratio objectives. This model transitions the focus from actions to responses, utilizing a history-conditioned sequence model to forecast future traffic volume and cost/value curves. It overcomes the shortcomings of control-based pacing and reinforcement learning techniques by simultaneously predicting horizon-aggregate results based on a single bid multiplier. The findings indicate that, under certain mild monotonicity conditions, the optimality gap compared to full per-tick control is limited by the variability of per-tick marginal value-per-cost. This method provides a simple analytical solution for non-stationary auction dynamics.