Transformers Are Inherently Succinct (2025)

Source: Hacker News

Abstract

We propose succinctness as a measure of the expressive power of a transformer in describing a concept. To this end, we prove that transformers are highly expressive in that they can represent formal languages substantially more succinctly than standard representations of formal languages like finite automata and Linear Temporal Logic (LTL) formulas. As a by‑product of this expressivity, we show that verifying properties of transformers is provably intractable (i.e., EXPSPACE‑complete).

Subjects

• Formal Languages and Automata Theory (cs.FL)
• Machine Learning (cs.LG)
• Logic in Computer Science (cs.LO)

Citation

• arXiv: arXiv:2510.19315 (cs.FL)
• Version: arXiv:2510.19315v2 (cs.FL)
• DOI: 10.48550/arXiv.2510.19315 (arXiv‑issued DOI via DataCite)

Submission history

• v1: Wed, 22 Oct 2025 07:25:54 UTC (28 KB) – submitted by Pascal Bergsträßer
• v2: Thu, 23 Oct 2025 08:09:19 UTC (28 KB)

View PDF | HTML (experimental)