AI That Buys Ads Just Got Smarter — More Clicks, Less Waste

A new paper adapts a training method called Group Relative Policy Optimization (GRPO) to improve automated bidding in online advertising. In this setting, advertisers hand over sequential bid decisions to algorithms that aim to maximize campaign value within constraints like a limited budget and a target cost-per-click. The authors say standard actor-critic reinforcement learning can be unstable, so they use GRPO—a critic-free method originally developed for large language model post-training—to fine-tune a strong heuristic baseline. They tested it on the BAT, iPinYou, and AuctionNet benchmarks. According to the article, the approach consistently outperforms baseline methods in clicks and is the best or second-best method in conversion volume.