NeurIPS 2024 Auto‑Bidding in Large‑Scale Auctions: Kuaishou Team Wins Both General and AIGB Tracks

During NeurIPS 2024, the Auto‑Bidding in Large‑Scale Auctions competition received 15,671 valid paper submissions (25.8% acceptance) and attracted over 1,500 participating teams from leading universities and companies.

The contest comprised two tracks: the General track, which focused on robust bidding under uncertain environments, and the AIGB (AI‑Generated Bidding) track, which required participants to model bidding as a generative sequence decision problem using large‑scale generative models.

In the General track, Kuaishou’s team applied a reinforcement‑learning‑based online exploration framework. They first solved an offline constrained optimization problem to obtain optimal bidding coefficients, then built a bidding simulator to learn long‑term value of bidding actions, and finally sampled within a coefficient interval to select the highest‑value action for online deployment.

For the AIGB track, the team introduced a Decision Transformer with RTG‑driven exploration. The method predicts the next return‑to‑go (RTG), generates exploratory bidding coefficients, and selects the better of the original and explored coefficients, mitigating out‑of‑distribution issues while encouraging higher‑RTG actions.

The competition task was a simplified Target CPA problem: given a budget B and target CPA C, the bidding agent must bid on N impression opportunities, maximize total conversions while keeping the realized CPA ≤ C. The environment used a Generalized Second Price (GSP) auction with three ad slots, where winning slots and payments depended on both bid values and estimated conversion probabilities.

Kuaishou’s solutions demonstrated a revenue increase of over 5% in their production advertising system, confirming the practical impact of reinforcement learning and generative‑model‑based bidding strategies.

The article also outlines the evolution of Kuaishou’s bidding algorithms from PID control to Model Predictive Control (MPC) and finally to reinforcement learning, highlighting the advantages of RL in handling uncertainty, multi‑step decision making, and long‑term optimization.

Looking ahead, the team plans to further explore the combination of reinforcement learning and generative models, as well as Monte‑Carlo Tree Search (MCTS) techniques, to advance the next generation of bidding algorithms.

At the end of the report, Kuaishou’s Commercial Algorithm team announced recruitment opportunities, inviting candidates from top universities and industry to join their AI‑driven advertising research group.