Anatomy of Associative Recall in Fixed-State Recurrences: A Matched-State Decomposition, an Interference Wall, and a Curriculum That Breaks It
Julian Boesch and Andrew Wee (Purdue University, Obit Research) decompose why fixed-state recurrent models lag on associative recall and find two cheap fixes.
Ask this paper
Convolution dominates: A short causal convolution adds about 0.5 recall in both families, so many architecture comparisons measure the missing convolution.
No class claim: The rank-1 delta rule's lead over diagonal transitions shrinks to 0.03 once both have the convolution, and a state-matched Mamba-2 ties it.
Interference, not capacity: Cells that solve 32-pair recall fall to chance with 4 pairs in a distractor haystack; a distance curriculum takes the same architecture from 0.021 to 1.000.
Curriculum control: Seed lock-in rises from 1/10 to 7/10, and gating the ramp on measured accuracy locks in 6/6 runs at L=512.
Abstract
Fixed-state recurrences--linear attention and state-space models--are reported to lag behind attention on associative recall, but whole-architecture comparisons cannot say which ingredient is responsible. We decompose masked multi-query recall at a fixed state budget along three single-knob axes: a short causal convolution, the transition structure (rank-1 delta rule vs. diagonal), and decay. The convolution dominates (~+0.5 recall in both families under matched training): comparisons that pit convolution-free cells against a convolution-equipped Mamba measure the missing convolution, not the recurrence. The rank-1 transition beats its diagonal ablation by +0.19/+0.32 at 16/32 pairs, but the margin shrinks to +0.03 once both cells carry the convolution, and a state-matched Mamba-2 ties the unarmed rank-1 cell: no class claim survives. Cells that solve 32-pair recall degrade gracefully with load yet fall to chance retrieving 4 pairs from a distractor haystack--flat across lengths and transitions. Interference under sparse supervision, not capacity: a distance curriculum takes the unchanged architecture from 0.021 to 1.000. Training is a lock-in lottery--a seed either locks in or does not--and the curriculum is the lever. Lock-in rises from 1/10 to 7/10 (p=0.02); dense supervision adds nothing; at L=256 a shaped ramp reopens a boundary the uniform curriculum cannot (4/5 vs. 0/9); and at L=512, where the ramp collapses (0/6), gating it on measured accuracy locks in 6/6 (p=0.001). Bidirectional denoiser cells, reading the query before the haystack, show no measurable advantage over causal training (ten seeds), and collision-key retrieval needs two layers. Arming for recall is free on an S_5 state-tracking guardrail--the armed cell is significantly better at every depth (p<=0.0044). These replace "recurrent models are bad at recall" with a measured decomposition and two cheap interventions.