Efficient Lookahead Encoding and Abstracted Width for Learning General Policies

arXiv:2605.18674v1 presents a new approach to generalized planning that improves upon Iterated Width (IW) policies. The authors address two key limitations of IW: unscalable compute costs and expressivity limitations when evaluating transitions individually, and inefficiency with large numbers of objects (e.g., IPC 2023 benchmarks). They introduce a holistic encoding of the entire search tree that jointly represents multiple transitions, and an abstracted width mechanism that reduces complexity by focusing on relevant atoms. The method achieves linear scaling with the number of objects while maintaining or improving policy quality. Experiments on IPC domains show significant speedups and better generalization compared to prior GNN-based approaches.