JKU Linz Opens KronLSTM With a Kronecker Delta Rule That Cuts Long-Context Attention Interference

Researchers at Johannes Kepler University Linz published the Kronecker Delta Rule on Oct 20, a doubly-gated update for associative matrix memories that corrects scalar associations along a single key-value direction rather than across all overlapping keys. Embedded in a KronLSTM architecture at 1.5B parameters and 100B training tokens, the model tops GDN-2 on five of six RULER recall tasks when extended to 32K context, hits 97.5-99.6% on the S-3 needle-in-haystack suite, and runs 11-18% faster than GDN-2 on an H100. The paper proves bounded memory norms and geometric convergence to the least-squares solution.