Gated DeltaNet-2 accepted to NeurIPS 2026, claims new SOTA for hybrid linear attention

AccBalanced · x · 2026-09-25

Gated DeltaNet-2 by Ali Hatamizadeh et al. has been accepted to NeurIPS 2026, and the authors report it as the new SOTA for hybrid/linear attention models.

Core idea: Linear attention replaces softmax attention's unbounded cache with a fixed-size recurrent state, but the hard part is editing this compressed memory without scrambling existing associations. Delta-rule models (Gated DeltaNet, Kimi Delta Attention/KDA) use a single scalar gate to control two different things: how much old content to erase and how much new content to commit.

Method: Gated DeltaNet-2 decouples these roles with a channel-wise erase gate bt and a channel-wise write gate wt, inheriting adaptive forgetting and channel-wise decay. It reduces to KDA when both gates collapse to the same scalar, and to Gated DeltaNet when decay also collapses. The paper derives a fast-weight update view, a chunkwise WY algorithm, and a gate-aware backward pass preserving efficient parallel training.

Results: At 1.3B parameters trained on 100B FineWeb-Edu tokens, it achieves the strongest overall results against Mamba-2, Gated DeltaNet, KDA, and Mamba-3 variants across language modeling and more. Paper and code are public.

Original post →

More from Models

Models channel →