PlatformBid benchmark rethinks auto-bidding for unified ad platforms

Real-time bidding in computational advertising traditionally features three players: Supply Side Platforms (SSP) selling impressions, Demand Side Platforms (DSP) bidding for advertisers, and Ad Exchanges running auctions. However, modern mega-platforms like social media and e-commerce companies now integrate all these functions internally. For these unified platforms, auto-bidding algorithms must simultaneously maximize advertiser conversions and total platform revenue—a dual objective that existing DSP-centric benchmarks fail to capture. To address this gap, a research team led by Shengtian Yang and seven co-authors introduced PlatformBid, the first comprehensive benchmark designed from a unified ad platform's perspective.

PlatformBid defines three representative settings that reflect real-world complexities: homogeneous competition (identical algorithms across advertisers), heterogeneous competition (diverse algorithmic strategies), and promotional competition (budget surges during events like Black Friday). The benchmark systematically evaluates a broad spectrum of existing auto-bidding methods, including classical control approaches, reinforcement learning (RL) methods, and emerging generative techniques. Beyond the benchmark, the authors propose BidFlow, a novel auto-bidding algorithm leveraging flow-matching—a generative modeling technique that provides expressive policy representation for dynamic competitive environments. Online experiments on Kuaishou, a major Chinese short-video platform, demonstrated a +0.68% improvement in target cost, offering deployment evidence and confirming PlatformBid's offline-online consistency as a reliable research foundation.