JD-BP: A Joint-Decision Generative Framework for Auto-Bidding and Pricing

A research team from JD.com and collaborating institutions has introduced JD-BP, a novel generative AI framework designed to solve a critical flaw in real-time advertising auctions. Current auto-bidding services use machine learning to set bids for advertisers aiming for goals like target return on investment (ROI). However, these systems suffer from 'ex-post' inefficiency—the optimal bid in hindsight differs from the real-time decision due to model prediction errors and system feedback latency. JD-BP's breakthrough is a joint architecture that doesn't just output a bid; it simultaneously generates a 'pricing correction term.' This additive correction works with auction payment rules (like the Generalized Second-Price, or GSP, mechanism) to adjust the final cost, effectively allowing the system to course-correct for past mistakes and uncertainties in real-time.

The framework employs several advanced techniques to make this work. It uses a 'memory-less Return-to-Go' design to encourage future value maximization without being crippled by historical constraint violations. A trajectory augmentation algorithm allows it to generate training data from any existing bidding policy, enabling easy 'plug-and-play' deployment. For training, the team combined a cross-attention module with an Energy-Based Direct Preference Optimization (DPO) method to refine the joint learning of bidding and correction. The results are significant: offline tests on the AuctionNet dataset showed state-of-the-art performance, but the real proof came from live A/B tests on JD.com's massive advertising platform. The online deployment confirmed a 4.70% lift in ad platform revenue and a 6.48% improvement in hitting target cost goals for advertisers.