Research map / Multi-agent systems and games / Human-AI and goal-conditioned
AuctionNet: A Novel Benchmark for Decision-Making in Large-Scale Games
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents AuctionNet, a benchmark for auto-bidding in large-scale ad auctions combining a deep-generative ad opportunity environment (LDM-based), a 500M-record pre-generated dataset, and baseline evaluations (Online LP, IQL, BC, DT, PID, Abid).
Problem
Researchers lack access to realistic large-scale ad auction game environments for bid decision-making; existing simulators (AuctionGym, AdCraft) ignore budget dynamics or model competitors unrealistically.
Approach
Build a data-driven ad auction environment in which ad opportunity features are generated by a latent diffusion model and values are predicted by an attention-based module, combined with a GSP auction module and diverse pre-trained auto-bidding agents, plus a released dataset and baseline evaluations.
Key results
- Online LP achieves the best relative performance among baselines in the basic task, with IQL and BC performing worse.
- Generated value-prediction MSEs are below the original standard deviations for all six tested quantities (e.g., value_time MSE 0.00176 vs std 0.00625).
- The dataset contains 10 million ad opportunities, 48 auto-bidding agents, and over 500 million auction records (80 GB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Online LP for auto-bidding instead of the evaluated baseline algorithms | AuctionNet auto-bidding · Basic task | mean episode reward | Helped |
| Apply the Target CPA constraint | all evaluated baseline algorithms · Target CPA task | reward | Hurt |
Surprising findings
- Online LP, a simple linear programming method, outperforms the RL and generative baselines (IQL, BC, DT) in the basic task.
- The generative model fails to capture the level-8 VIP drop in consumption amount, attributed to rarity of that data.
Limitations
- Generated data show biases in some details, such as specific VIP levels and categories with many values.
- Generative model performance could be improved.
- Dataset release is deferred until after the competition.
Experimental setup
Datasets: real-world Taobao/Alibaba online advertising data
Benchmarks: AuctionNet basic task (BCB) AuctionNet Target CPA task
Metrics: normalized mean episode reward relative to Abid MSE between generated and original value distributions 3D PCA visual similarity
Baselines: PID Controller Online LP IQL Behavior Cloning Decision Transformer Abid
Builds on: Latent Diffusion Model Generalized Second Price auction Decision Transformer IQL AuctionGym
Closest papers by meaning
- Auctionformer: A Unified Deep Learning Algorithm for Solving Equilibrium Strategies in Auction Games (ICML 2024 · similarity 0.88)
- LLM-Auction: Generative Auction towards LLM-Native Advertising (NeurIPS 2026 · similarity 0.88)
- Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms (AAAI 2024 · similarity 0.88)
- Increasing Revenue in Efficient Combinatorial Auctions by Learning to Generate Artificial Competition (AAAI 2025 · similarity 0.88)
- GenAuction: A Generative Auction for Online Advertising (AAAI 2025 · similarity 0.87)
- Optimal Auction Design in the Joint Advertising (ICML 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.