The Illusion of State in State-Space Models

William Merrill (NYU), Jackson Petty (NYU), Ashish Sabharwal (Allen Institute for AI) · ICML 2024
Does a recurrent form buy recurrent power?
arXiv:2404.08819

Where the models land assumes TC0 ≠ NC1

Conceptual map of complexity classes. Click a chip. Toggle the conjecture.
Click a chip to see where it sits.

Three assumptions behind the claim

Five people, five coins

All dimes except Carl's coin (orange). Swap neighbours and track the odd coin. Final state is one of 120 arrangements: S5.
No swaps yet.

Order matters: swaps don't commute

Same two swaps, opposite order, different arrangement (computed live from the identity).

Group ladder: size of each group

Group orders are exact. Barrington: word problem of non-solvable groups (smallest: S5) is NC1-complete.

Why linear recurrences parallelize

hi = Ai·hi-1 + Bi·xi: no nonlinearity inside, so the combine step is associative. Step through 8 tokens.

Depth vs. sequence length

Sequential steps (n−1) vs. a balanced tree (⌈log₂ n⌉). Exact arithmetic, not measurements.

Proof sketch: why S4 and Mamba stay in TC0

Click any box. Theorem numbers follow the episode.
Click a node.

Transition matrices per model Illustrative values

Structure only: S4 reuses one matrix, Mamba varies a diagonal, IDS4 varies a full matrix.

IDS4 fix: automaton as matrix product

Each symbol becomes a matrix; multiplying composes state maps. Shown: parity automaton from a one-hot start vector set by the beginning-of-string symbol.

Layers needed for >90% accuracy vs. length Illustrative

Reported: RNN and IDS4 need 1 layer for lengths 5–20 on all three groups. Transformer, S4, Mamba climb from ≈1 to ≈4 (read off the plot; no exact counts, no variance). Dashed lines join those two readings only.

The three 60-element groups

What the experiment does and doesn't show

References