Kuaishou's KDD 2026 Paper: Introducing PlatformBid, First Platform-Perspective Auto-Bidding Benchmark
机器之心 · wechat · 2026-08-04
Kuaishou, collaborating with Southeast University and NTU, introduced PlatformBid, the industry's first auto-bidding benchmark designed from a unified advertising platform's perspective, along with BidFlow, a new method based on Flow Matching. The research has been accepted by KDD 2026.
While traditional auto-bidding research focuses on individual advertisers (DSP perspective), PlatformBid incorporates platform-level constraints and introduces three evaluation settings: homogeneous, heterogeneous, and promotional competition. This accurately reflects the dynamic game theory and global efficiency of multiple advertisers in real-world scenarios.
To address the challenges of multimodal bidding distributions, the proposed BidFlow algorithm leverages Flow Matching for distribution modeling and utilizes Q-value guided distillation for efficient single-step inference. Experiments show that BidFlow achieves SOTA in most settings and has been fully deployed in Kuaishou's e-commerce advertising system, demonstrating strong offline-to-online consistency.
More from Research
- GPTQ-2D: Cubic-Time Two-Sided Adaptive Rounding for Matrix Quantization — ISTA-DASLab · 2026-08-04
- DreamTraj: Predicting 6-DoF Trajectories by Reading Unrendered Video Diffusion Latents — SUSTech · 2026-08-04
- GradCuit: Credit-Assigned Gradient Flow for Robust Test-Time Latent Reasoning — Zhaoxin Yu · 2026-08-04
- Kuaishou's RecHarness: Automating Recommender Optimization with Bandit-Routed LLM Agents — kuaishou · 2026-08-04
- GUSH3R: Single Forward Pass for Dynamic Human-Scene 3D Reconstruction — rsasaki0109 · 2026-08-04
- MIT Study: Non-Experts Trust Wrong AI Diagnoses, Clinicians Catch Errors — nordicinst · 2026-08-04