CLMay 29, 2025

Characterizing the Expressivity of Transformer Language Models

arXiv:2505.23623v19 citationsh-index: 10
Originality Incremental advance
AI Analysis

This work provides a foundational theoretical framework for understanding transformer expressivity, which is incremental but crucial for the machine learning and AI community.

The paper tackled the problem of theoretically characterizing the expressive power of transformer language models under realistic assumptions like fixed precision and soft attention, showing they are exactly as expressive as a specific fragment of linear temporal logic and providing empirical validation of this capacity.

Transformer-based language models (LMs) have achieved widespread empirical success, but their theoretical expressive power remains only partially understood. Prior work often relies on idealized models with assumptions -- such as arbitrary numerical precision and hard attention -- that diverge from real-world transformers. In this work, we provide an exact characterization of fixed-precision transformers with strict future masking and soft attention, an idealization that more closely mirrors practical implementations. We show that these models are precisely as expressive as a specific fragment of linear temporal logic that includes only a single temporal operator: the past operator. We further relate this logic to established classes in formal language theory, automata theory, and algebra, yielding a rich and unified theoretical framework for understanding transformer expressivity. Finally, we present empirical results that align closely with our theory: transformers trained on languages within their theoretical capacity generalize perfectly over lengths, while they consistently fail to generalize on languages beyond it.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes