✳flâneur — a map of the web's best reading

[2603.19954] On the Ability of Transformers to Verify Plans

arxiv.org · 650 words · saved by 1 readers

Abstract: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.

[2603.19954] On the Ability of Transformers to Verify Plans Skip to main content arXiv is now an independent nonprofit! Learn more × Search arXiv Press Enter to search · Advanced search --> Computer Science > Artificial Intelligence arXiv:2603.19954 (cs) [Submitted on 20 Mar 2026 ( v1 ), last revised 3 Jul 2026 (this version, v2)] Title: On the Ability of Transformers to Verify Plans Authors: Yash Sarrof , Yupei Du , Katharina Stein , Alexander Koller , Sylvie Thiébaux , Michael Hahn View a PDF of the paper titled On the Ability of Transformers to Verify Plans, by Yash Sarrof and

Explore this link on the map →

related reading