A new bidding algorithm learns from bad past decisions instead of just copying them, and it still moved the needle in a live test.
Researchers built QGA, a Q-value regularized generative auto-bidding method that plugs a double Q-learning value check into a Decision Transformer, the neural-network architecture usually used for sequence-based reinforcement learning. Auto-bidding software sets how much an advertiser bids in real time, and most current systems learn by imitating historical bidding logs. The catch is that those logs are full of suboptimal decisions, so a model that just copies them inherits the mistakes. QGA's fix scores every candidate bid with the Q-value module during training and exploration, steering the policy toward actions with better estimated payoff instead of just the ones that happened to appear in the data.
This targets a real weak spot in the current wave of generative bidding systems, which behave like sophisticated autocomplete for auction decisions and can plateau exactly where their training data plateaus. In large-scale real-world A/B testing, the researchers report a 3.27% increase in ad GMV and a 2.49% improvement in ad ROI - modest on paper, but at ad-marketplace scale, single-digit percentage points move real budgets.
Auto-bidding papers routinely claim wins like this against their own simulation environments and unnamed baselines, so the honest read here is a solid incremental fix, not the auction-clearing breakthrough the framing suggests.