Why Transformers Outperform RNNs: Polynomial Parameters Compact vs Exponential for Certain Languages

chrmanning · x · 2026-08-19

Chris Manning highlights the paper 'Transformers are Inherently Succinct', offering a theoretical reason for Transformers' practical superiority over RNNs. The study shows that fixed-precision polynomial-size transformers can compactly represent certain languages that would require an exponential number of parameters for RNNs to represent.

Related event: New Paper: Transformers Are Inherently Succinct, Explaining Edge Over RNNs(2 posts)→

Original post →

More from Research

Research channel →