On the Ability of Transformers to Verify Plans
This addresses the theoretical gap in transformer generalization for AI planning, which is important for researchers in machine learning and AI planning.
The paper tackles the problem of understanding when transformers can generalize in AI planning tasks by analyzing their ability to verify plans, introducing C*-RASP to prove length generalization guarantees for a large class of planning domains.
Transformers have shown inconsistent success in AI planning tasks, and theoretical understanding of when generalization should be expected has been limited. We take important steps towards addressing this gap by analyzing the ability of decoder-only models to verify whether a given plan correctly solves a given planning instance. To analyse the general setting where the number of objects -- and thus the effective input alphabet -- grows at test time, we introduce C*-RASP, an extension of C-RASP designed to establish length generalization guarantees for transformers under the simultaneous growth in sequence length and vocabulary size. Our results identify a large class of classical planning domains for which transformers can provably learn to verify long plans, and structural properties that significantly affects the learnability of length generalizable solutions. Empirical experiments corroborate our theory.