Auto‑Bidding in Large‑Scale Auctions: The AIGB Paradigm and the AuctionNet Benchmark

At NeurIPS 2024, Alibaba’s advertising platform (AliMama) organized a special workshop titled “Auto‑Bidding in Large‑Scale Auctions,” which was also selected as a competition track. The workshop was jointly proposed by Chinese researchers from Peking University and Alibaba.

AIGB (AI‑Generated Bidding) is introduced as a new solution for the auto‑bidding problem that leverages generative models such as Transformers and Diffusion Models. Unlike traditional reinforcement‑learning (RL) approaches, AIGB treats bidding as a strategy‑generation task, avoiding value‑function estimation errors and handling long‑sequence sparse‑reward scenarios more effectively.

Alibaba released the world‑first large‑scale benchmark “AuctionNet,” containing 48 interacting agents, over 5 billion bidding records (≈80 GB). The benchmark was highlighted in a NeurIPS 2024 Spotlight paper and is intended for research on both advertising bidding and broader large‑scale game decision‑making.

The competition featured two tracks – an AIGB track (generative‑model agents) and a General track (bidding under uncertainty). More than 1 861 participants formed 1 522 teams (including 948 in the AIGB track). Winners such as the KGAB and CleanDiffuser teams received cash prizes and internship opportunities at Alibaba.

AIGB Track: Learn auto‑bidding agents via generative models.

General Track: Auto‑bidding under uncertainty.

“AuctionNet not only supports research on advertising bidding algorithms but also serves as a standard benchmark for decision‑making in large‑scale games,” a Alibaba technical lead remarked.

The benchmark and the AIGB paradigm have implications beyond online advertising, extending to games, autonomous driving, recommendation systems, and quantitative finance. While the field still faces challenges such as sparse conversions and complex auction mechanisms, open datasets like AuctionNet aim to foster collaboration and accelerate progress.