Technical Summary: Constrained Auto-Bidding via Generative Response Modeling

1. Problem Statement

Online advertising platforms face a complex constrained optimization problem: maximizing advertiser value (e.g., conversions or revenue) over a long campaign horizon while adhering to strict budget limits and efficiency targets (such as Cost-Per-Acquisition, CPA, or Return-On-Ad-Spend, ROAS).

The core challenges identified in the paper are:

• Non-stationarity and Uncertainty: Future traffic volumes and auction dynamics are uncertain and change over time.
• Coupled Constraints: Budget and efficiency constraints span the entire horizon, yet bidding decisions are made impression-by-impression (or tick-by-tick).
• Limitations of Existing Approaches:
  • Control-based pacing: Reacts to deviations (e.g., overspending) but cannot anticipate future conditions.
  • Reinforcement Learning (RL) & Generative Models: Often fold constraints into reward signals (e.g., Return-to-Go), which obscures constraint violations and leads to brittle performance under distribution shifts.
  • Bid-landscape models: Predict per-auction win rates but treat auctions as snapshots, lacking a mechanism to enforce horizon-level constraints directly.

The authors argue that existing methods struggle to balance anticipation of future conditions with explicit, transparent constraint enforcement.

2. Methodology: Generative Response Model (GRM)

The paper proposes a paradigm shift: instead of learning the optimal action (the bid multiplier), the system learns the environment's response to a given action.

2.1 Core Concept: Response Prediction

The Generative Response Model (GRM) is a history-conditioned sequence model. Rather than predicting a single bid multiplier for the current tick, it predicts a horizon-aggregate response bundle conditioned on the pre-decision history $H_t$. This bundle consists of:

1. Future Traffic Volume ($\hat{I}_{t:T}$): The total number of bid opportunities remaining in the horizon.
2. Aggregate Cost Curve ($\hat{\bar{C}}_{t:T}(\alpha)$): The expected cost per opportunity as a function of a single bid multiplier $\alpha$ over the remaining horizon.
3. Aggregate Value Curve ($\hat{\bar{V}}_{t:T}(\alpha)$): The expected value per opportunity as a function of $\alpha$.

2.2 Model Architecture and Training

• Architecture: GRM utilizes a causal Transformer encoder to process the history of contextual features ($s_{1:t}$) and executed multipliers ($\alpha_{1:t-1}$). A decoder outputs a low-dimensional parameter vector that defines the traffic forecast and the parameters for the cost and value curves.
• Curve Parameterization: The cost and value curves are parameterized using a log-sigmoid family (based on the normal CDF). This choice reflects economic realities:
  • Saturation at high multipliers (diminishing returns).
  • Monotonicity (higher multipliers yield more spend/value).
  • Boundary conditions (zero spend at $\alpha=0$).
• Training (Future-Sampling Supervision): Since offline data only contains outcomes for logged multipliers (not counterfactuals), GRM is trained using a future-sampling strategy. For an anchor tick $t$, the model samples future ticks $k$ from the logged data and fits the predicted aggregate curve to the realized outcomes at those future ticks. The loss function is a traffic-weighted regression, ensuring high-traffic periods contribute more to the curve fitting.

2.3 Analytic Min-Pacing Controller

Once GRM predicts the response curves, a lightweight, analytic controller determines the bid multiplier $\alpha_t$ for the current tick. This controller does not learn; it solves for feasibility:

1. Budget Pacing ($\alpha_B$): Solves for $\alpha$ such that the predicted total cost equals the remaining budget ($\hat{C}_{t:T}(\alpha) = B_t$).
2. CPA Pacing ($\alpha_C$): Solves for $\alpha$ such that the predicted total cost satisfies the efficiency constraint ($\hat{C}_{t:T}(\alpha) - \tau \hat{V}_{t:T}(\alpha) = \Delta_t$).
3. Execution: The final multiplier is the conservative minimum: $\alpha_t = \min\{\alpha_B, \alpha_C\}$.

This process is repeated at every tick (receding horizon), updating the state and re-solving based on the latest predictions.

3. Key Contributions and Theoretical Results

The paper provides three main theoretical guarantees that link prediction quality to constraint satisfaction:

1. Single-$\alpha$ Approximation Gap (Theorem 5.2): The authors prove that the gap between the optimal solution using a single multiplier for the whole horizon versus a per-tick optimal solution is bounded by the efficiency dispersion (the variance of per-tick marginal value-per-cost). If efficiency is relatively uniform, the single-curve approximation is near-optimal.
2. Exactness of Min-Pacing (Theorem 5.3): Under mild monotonicity assumptions, the min-pacing rule ($\min\{\alpha_B, \alpha_C\}$) is proven to be the exact solution to the single-multiplier constrained optimization problem.
3. Constraint Violation Bounds (Theorem 5.4): The paper proves that under receding-horizon replanning, constraint violations scale linearly with prediction error. Specifically, the total budget and CPA violations are bounded by terms involving the curve prediction errors ($\epsilon_C, \epsilon_V$) and traffic prediction error ($\epsilon_I$). This establishes a "graceful degradation" property: better predictions directly lead to safer constraint adherence.

4. Experimental Results

Experiments were conducted using the AuctionNet simulation environment (from the NeurIPS 2024 Auto-Bidding Challenge).

• Performance: GRM achieved an average score of 33.88, outperforming the strongest baseline (EBaReT, a generative approach) by 7.8%. It also significantly outperformed standard offline RL methods (CQL, IQL) and Behavioral Cloning.
• Robustness to Distribution Shift:
  • Competition Surge: When competitor budgets increased by 1.1x, GRM's performance degraded by only 7.2%, compared to 22.6% for the Decision Transformer (DT) baseline.
  • CPA Tightening: When the target CPA was tightened to 0.8x, GRM degraded by 5.0%, while DT degraded by 13.9%.
  • The authors attribute this robustness to GRM's ability to re-predict response curves at every tick, allowing it to adapt to the changed competitive landscape, whereas DT relies on static offline distributions.
• Prediction-Performance Correlation: A strong negative correlation ($r = -0.78$) was observed between the model's validation loss and its test score. This empirically validates Theorem 5.4, confirming that improvements in prediction quality directly translate to better bidding performance.
• Ablation Studies: The log-sigmoid curve family was found to be superior to linear, piecewise-linear, and standard sigmoid parameterizations, as it correctly captures the saturation and diminishing returns inherent in bidding landscapes.

5. Significance and Claims

The paper claims significance in shifting the learning objective from actions to responses. By predicting the environment's aggregate response curves rather than the optimal action itself, the authors achieve:

• Explicit Constraint Handling: Constraints are enforced via analytic root-finding on predicted curves, making constraint violations transparent and traceable to specific prediction errors. This contrasts with RL methods where constraints are often hidden within reward shaping.
• Stability under Shift: The separation of prediction (GRM) and control (Min-Pacing) allows the system to adapt to distribution shifts more effectively than end-to-end generative models that rely on fixed conditioning signals.
• Theoretical Grounding: The work provides formal bounds on how prediction errors propagate to constraint violations, offering a theoretical guarantee of "graceful degradation" that is often missing in deep learning-based bidding systems.

The authors position GRM as a practical, theoretically grounded approach that bridges the gap between predictive modeling and constrained optimization in high-stakes, non-stationary environments like real-time advertising.