Gated DeltaNet-2

Summary

Gated DeltaNet-2 is the linear recurrent attention method introduced by Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention. It replaces the scalar delta-rule update gate used by Gated DeltaNet and KDA with separate channel-wise key-side erase and value-side write gates.

Role In The Wiki

Use this page as the object card for the Gated DeltaNet-2 method. The source page carries the detailed evidence, limitations, and agenda mapping.

Gated DeltaNet-2 is relevant to the wiki as upstream architecture evidence for compact recurrent state and memory editing. It is not direct evidence for numeric time-series modeling, event streams, or action-conditioned world models.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in gated-deltanet-2-2026 rather than duplicating verdict rows here. At the entity level, the important transfer is the state-update pattern: erasing stale associations and writing new information may need separate gates when a fixed-size latent state must preserve long-context information under interference.

Evidence

• Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention

Official Artifacts

• Preprint: arXiv 2605.22791
• Official NVIDIA page: Gated DeltaNet-2 publication page
• Official code: NVlabs/GatedDeltaNet-2 — GitHub API reported 221 stars, 17 forks, and SPDX NOASSERTION; the project declares NVIDIA Source Code License-NC in pyproject.toml and LICENSE at ingest time.
• Official X announcement: Ali Hatamizadeh post.
• Official released models: none found at ingest time.

Related Pages

• Efficient Recurrent Sequence Models
• Streaming Latent-State Updates
• Time-Series Scaling And Efficiency
• Foundation Time-Series Model Research Agenda
• Mamba-3
• Mamba-2
• RWKV-TS
• LT2