SSMs' Expressive Power Limited: Mamba-style Models Fail at State Tracking, Paper Shows

davidbau · x · 2026-08-21

In a discussion on AI architectures, davidbau cites a paper by Merrill et al., 'The Illusion of State in State-Space Models,' which shows that state-space models (SSMs) have similar expressive limitations to transformers, confined to TC0, and cannot solve simple state-tracking problems like permutation composition, thus failing to track chess moves, evaluate code, or track entities in long narratives. Experiments confirm Mamba-style SSMs struggle with state tracking. DeltaNet is mentioned as a potential alternative.

Original post →

More from Models

Models channel →