CATALIST: Structurally Guided Sequence Learning withModel-Based Augmentation for Offline Auto-Bidding

Abstract Offline auto-bidding learns bidding policies from logged auction trajectories to optimize resourceallocation under budget and cost-per-acquisition (CPA) constraints. Temporal dependencies amongauction features and limited action coverage in the logs complicate policy learning. This paper pro-poses CATALIST, a structurally guided sequence-learning framework that combines dependencypriors with local model-based augmentation. CATALIST constructs structured representations usinglagged and instantaneous dependency priors and models decision histories through token-type andtemporal attention biases. Alternative actions are generated within the same historical context andselected through resource-feasibility checks and model-predicted gain screening for action supervi-sion. Training jointly incorporates action learning from observed and generated samples, outcome andvalue prediction using only observed records, and context-consistency regularization. On AuctionNet,CATALIST achieves a Score of 260.95 at the nominal budget, improving over CDT’s 175.48 by 48.7%.At the 150% budget ratio, it achieves 339.24, exceeding GAS and GAVE, which obtain 294.13 and304.32, by 15.3% and 11.5%, respectively. The experimental results show that combining structurally guided sequence learning with model-based augmentation improves offline auto-bidding performance.