Why Auto‑Bidding in Large‑Scale Auctions Is the Hottest NeurIPS Challenge

Auto‑Bidding in Large‑Scale Auctions

When a user searches for a product on an e‑commerce platform, the system launches an instantaneous ad auction. Advertisers submit bids and an automatic bidding engine computes the optimal bid by jointly considering user profile, behavior data, advertiser goals, budget constraints, and the current auction environment. The highest‑scoring ad is displayed together with organic results, all within a few milliseconds.

Four Generations of Auto‑Bidding Strategies

First Generation – Classic Control : Transform ROI maximization into a budget‑consumption control problem. PID controllers and similar algorithms regulate spend along a predefined curve.

Second Generation – Planning / Linear Programming : Predict next‑day traffic, formulate a linear program to solve for optimal bid parameters, and re‑optimize online as new data arrives. Accuracy depends on traffic forecasts.

Third Generation – Reinforcement Learning : Model bidding as a sequential decision‑making problem. Early methods use online RL; later work adopts offline RL to approximate the online data distribution, and the newest approaches employ online RL that interacts directly with the live auction environment.

Fourth Generation – Generative Models : Use large‑scale generative AI (e.g., diffusion models, Transformers) to model the joint distribution of bids, objectives, and constraints. Alibaba Mama’s AIGA (AI Generated Action) framework and its AIGB (AI Generated Bidding) variant exemplify this shift.

AIGB Research Initiative

The PAAI team (Peking University‑Alibaba Mama AI Innovation Joint Lab) identified error accumulation in long‑horizon RL bidding. They proposed AIGB, which treats bids, objectives, and constraints as a joint probability distribution and reframes the problem as conditional generation. This directly links decision trajectories with returns, reducing cumulative error. A related paper has been accepted at KDD 2024.

Competition Tracks

AIGB Track : Develop generative‑model‑based agents (e.g., diffusion models, Transformers, foundation models) that produce precise bid decisions over long sequences.

General Track : Address uncertainty in large‑scale auctions, handling stochastic user arrivals, conversion variance, data sparsity, and opponent strategy shifts.

Both tracks provide roughly 500 million auction data points and a full training framework, offering a substantial research resource for decision‑intelligence, reinforcement learning, game theory, and generative‑AI.

Participation Details

Registration is open on the Tianchi platform:

AIGB Track: https://tianchi.aliyun.com/competition/entrance/532236

General Track: https://tianchi.aliyun.com/competition/entrance/532226

Top performers can receive up to $6 000 in cash and internship or recruitment opportunities.