PlatformBid: An Auto-Bidding Benchmark from a Unified Advertising Platform's Perspective
Shengtian Yang, Yewen Li, Peng Jiang, Zhiyi Lyu, Bo An, Peng Jiang, Qingpeng Cai, Lei Feng
cs.LG
2026-07-29
First platform-view auto-bidding benchmark with three competition settings; its flow-matching method BidFlow tops two settings and lifts Kuaishou online target cost +0.68%.
Auto-bidding sits at the core of computational advertising: advertisers set a budget and a cost-per-action (CPA) ceiling, and an algorithm bids for them in real-time auctions. For over a decade, the research and the benchmarks, from the iPinYou dataset to recent work, have taken the DSP view, treating one advertiser as a lone agent whose only goal is to maximize its own conversions under a CPA constraint.
Meta, TikTok, and Kuaishou long ago fused SSP (selling impressions), DSP (bidding for advertisers), and Ad Exchange (running the auction) into a single system. From the platform's side, a good algorithm has to optimize total platform revenue and ecosystem health, not just advertiser conversions. A policy that looks optimal in a DSP-only benchmark can, once every advertiser runs it at once, drive a race to the bottom that erodes platform income. The gap between how auto-bidding is graded and how it is deployed has never been systematically closed.
PlatformBid does two things: it rewrites the grading rubric, and it ships a new method.
On the rubric. The benchmark builds on the AuctionNet dataset (48 advertisers, 480K training trajectories, Dense and Sparse reward variants), but the key shift is from isolated single-advertiser optimization to real-time interaction among many advertisers in a shared auction. A floor price stops advertisers from colluding on low bids that hurt platform revenue. Platform-level metrics (total conversion, budget utilization) sit beside advertiser-level ones (CPA ratio, exceed rate). The composite Score is the headline: count raw conversion when CPA is within target, but multiply conversion by (targetCPA / actualCPA)^2 once it slips over, folding cost discipline and conversion into one number.
Three competition settings:
The new method, BidFlow. The authors argue the competition settings make bidding distributions multimodal, which deterministic policies cannot capture. BidFlow trains a behavior-cloning policy with flow matching to model that multimodal distribution, but flow models need iterative sampling at inference and are slow. So it distills the flow policy into a one-step policy that keeps the expressiveness but bids in a single forward pass, guided toward high return by a critic's Q-values.
On the Dense dataset, composite Score:
| Setting | BidFlow | Best baseline | Note |
| Homogeneous | 348.04 | GAS 314.27 | CPA ratio 0.88, lowest exceed rate |
| Heterogeneous (target group) | 312.69 | DT score 321.96 | BidFlow loses this one |
| Promotional | 356.20 | GAS 333.84 | All-advertiser basis |
The online test is the most persuasive part. In a Kuaishou production A/B against a heavily tuned DT, BidFlow lifted impressions +0.27%, spend +0.30%, and target cost +0.68%. On the public iPinYou dataset it ranks first across all three settings.
One finding runs through every table: lower CPA-variance strategies score higher. Steady, cooperative bidding beats aggressive gaming for both platform and advertiser.
For ad-tech and RL practitioners the value is on two layers. The benchmark itself validates the offline-online gap with a +0.68% online lift, which means conclusions drawn here are likely to transfer to a real platform rather than just look good in a paper table. For anyone working on auctions or multi-agent RL, this is a usable multi-agent testbed.
The second layer is the BidFlow recipe: learn an expressive multimodal policy with flow matching, then distill it to single-step inference. That pattern travels beyond bidding to any offline RL task that wants a generative policy but is bottlenecked on inference speed.
The authors name three: the benchmark covers only bidding algorithms, not auction mechanism design; it does not model relationships among advertisers, and real platforms host millions of them, far beyond 48.
A few more caveats. BidFlow degrades sharply in heterogeneous competition with sparse rewards (target-group Score 35.97 against DT score's 43.91), which the authors blame on hard modeling under sparse feedback without offering a fix, and that happens to be the most common real-world condition. The 48-advertiser, 30-minute-update setup is still a step away from second-by-second, million-advertiser reality. The +0.68% online lift is measured against an already heavily tuned DT, so the absolute gain is modest, and only the relative target-cost number is reported.