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

2026-05-21

AI Summary

This work addresses the limitations of existing linear attention mechanisms in compressive memory editing, which struggle to balance forgetting and writing due to reliance on a single scalar gating signal that restricts expressivity. The authors propose Gated DeltaNet-2, the first linear attention architecture to decouple erase and write operations by introducing channel-wise independent erase and write gates, enabling fine-grained memory control. This approach unifies and generalizes Gated DeltaNet and KDA, integrating adaptive forgetting, channel-wise decay, and a block-wise WY algorithm, while introducing a gate-aware backpropagation mechanism that supports efficient parallel training and recurrent inference. Trained on 1.3B parameters and 100B FineWeb-Edu tokens, the model consistently outperforms Mamba-2/3, Gated DeltaNet, and KDA across language modeling, commonsense reasoning, and long-context multi-key retrieval (RULER) benchmarks.

Abstract

Linear attention replaces the unbounded cache of softmax attention with a fixed-size recurrent state, reducing sequence mixing to linear time and decoding to constant memory. The hard part is not just what to forget, but how to edit this compressed memory without scrambling existing associations. Delta-rule models subtract the current read before writing a new value, and Kimi Delta Attention (KDA) sharpens forgetting with channel-wise decay. But the active edit still uses a single scalar gate to control two different things: how much old content to erase on the key side and how much new content to commit on the value side. We introduce Gated DeltaNet-2, which generalizes both Gated DeltaNet and KDA by inheriting adaptive forgetting and channel-wise decay while addressing their shared limitation, the scalar tie between erasing and writing. Gated Delta Rule-2 separates these roles with a channel-wise erase gate b_t and a channel-wise write gate w_t, reducing to KDA when both gates collapse to the same scalar and to Gated DeltaNet when the decay also collapses. We derive a fast-weight update view, a chunkwise WY algorithm with channel-wise decay absorbed into asymmetric erase factors, and a gate-aware backward pass that preserves efficient parallel training. At 1.3B parameters trained on 100B FineWeb-Edu tokens, Gated DeltaNet-2 achieves the strongest overall results among Mamba-2, Gated DeltaNet, KDA, and Mamba-3 variants across language modeling, commonsense reasoning, and retrieval. Its advantage is most pronounced on long-context RULER needle-in-a-haystack benchmarks, where it improves the evaluated multi-key retrieval setting and remains strong in both recurrent and hybrid settings. Code is available at https://github.com/NVlabs/GatedDeltaNet-2.

Problem

• linear attention
• memory editing
• erase-write decoupling
• gated mechanisms
• recurrent state

Innovation

• linear attention
• memory editing
• erase-write decoupling
• channel-wise gating
• fast weight