Transformer Architecture · 2024

Gated Delta Networks: Improving Mamba2 with Delta Rule (Gated DeltaNet)

Songlin Yang, Jan Kautz, Ali Hatamizadeh (NVIDIA)

Unifies gating (fast memory erasure) with the delta rule (targeted memory updates) into one hardware-parallel linear-attention layer, becoming the recurrent primitive behind later hybrid models.

Editorial record

Plain-language summary

Unifies gating (fast memory erasure) with the delta rule (targeted memory updates) into one hardware-parallel linear-attention layer, becoming the recurrent primitive behind later hybrid models.

Knowledge graph

Relationships

Descendants

  • Depends onEvidence: Strongly supported

    Attention Is All You Need

    Builds on the transformer architecture lineage in the archive

    freshness sweep 2026

Source record

Provenance

Record ID
P-608
Record created
2026-07-13
Last reviewed
2026-07-14
Record version
2

Citation caveat: Citation metadata is approximate and marked unverified in the source dataset.