Dissecting AI

All papers

Kronecker Delta Rule for Long Context Linear Attention

Anamaria-Roberta Hartl, Levente Zólyomi, Michael List, Günter Klambauer, Sebastian Böck, Sepp Hochreiter

arXiv 2610.26834 53 pages Read the paper

Modules from this paper

Chunkwise-Parallel Execution of KronLSTM

Token-by-token KronLSTM updates serialize memory access; chunkwise execution turns the within-chunk dependencies into a Gram-based triangular solve and propagates memory only between chunks.

KronLSTMchunkwise-parallelismlinear-attentionrecurrent-memory

Gated KronLSTM Sequence Mixer

Recurrent linear attention needs controlled memory updates; KronLSTM separates forgetting, erasing, and writing into gates inside a standard sequence-mixer block.

linear-attentionsequence-modelinggatingrecurrent-memory

Pair-Targeted Kronecker Delta Memory Updates

Hebbian writes can bury an old association and ordinary delta updates can disturb overlapping pairs; the Kronecker Delta Rule corrects one pair’s scalar association along that pair’s rank-one direction.

linear-attentionassociative-memorymemory-updateskronecker-delta

Stable Expected Dynamics of Kronecker Delta Memory

When random associations cannot all be fit at once, the Kronecker Delta memory need not settle to a single state; its expected update follows least squares, while its stochastic state settles into a bounded stationary regime.

kronecker-deltamemorystabilityleast-squares