arXiv:2608.30310v1 Announce Type: cross Abstract: Hybrid large language models interleave full-attention layers with linear-attention layers to reduce the cost of long-context inference. This structu...
The paper investigates how block‑diffusion language models can use a constant‑size cache to enable efficient parallel decoding. By employing sequence mixers that summarize completed blocks into a reusable state and a block‑causal training objective, the authors pretrain three 3B block‑diffusion denoisers (attention, Mamba, and hybrid) on 300 B tokens. The resulting state‑space cache remains O(1) in memory and latency regardless of context length, yielding significant speed‑up and memory savings compared to traditional attention‑based caches, especially at very long sequences.
The paper introduces DAMP, a decay‑aware mixed‑precision quantization scheme for recurrent‑state representations in GDN and KDA language models. By identifying high‑risk channels through quantization‑error energy and decay persistence, DAMP stores these channels at higher precision while compressing the rest to INT8, achieving a 9.9‑bit average precision. Experiments on Qwen3.6‑35B and Kimi‑Linear‑48B show a 69.1% reduction in recurrent‑state storage, up to 2.01× faster state‑update kernels, and up to 10.9% lower full‑model TPOT while preserving accuracy close to the FP32 baseline.
DeltaS is a query‑agnostic, training‑free method for evicting key‑value cache entries in hybrid video‑language models that combine linear and full attention. It uses the change in the recurrent state of gated‑delta linear attention—called state drift—to decide which video chunks to keep, selecting those that induce larger normalized state changes. In experiments with a fixed memory budget, DeltaS outperforms position‑, attention‑, and key‑value‑based eviction signals, improving performance by 2.1 points on average across six long‑video benchmarks and 5.6 points on the longest benchmark, while adding only 1.9% of the forward‑pass cost.
ReplaySSM is a technique that redefines how state-space models (SSMs) handle their recurrent state by caching SSM inputs instead of the state itself. It unrolls the recurrence into a checkpoint plus a buffer of recent inputs, reducing state memory traffic by about half and enabling efficient speculative rollback. The method also allows separate precision levels for checkpoints, improving output accuracy while accelerating decoding in hybrid models such as Mamba-2 and Gated DeltaNet.
arXiv:2608. 15533v1 Announce Type: cross Abstract: Linear attention models eliminate the quadratic prefix computation and context-growing KV cache of softmax attention by replacing pairwise token interactions with recurrent state updates.
arXiv:2610.14754v1 Announce Type: new Abstract: Hybrid language models interleave Gated DeltaNet (GDN) layers, whose recurrent memory has a fixed size, with full-attention layers, whose key-value (KV...
arXiv:2602.08005v2 Announce Type: replace-cross Abstract: Efficient long-context inference faces two coupled bottlenecks: KV-cache memory grows linearly with context length, while attention computati...
arXiv:2609.33051v2 Announce Type: replace Abstract: Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches wh...
Minima-KV introduces a retention‑preserving hierarchy for mixed‑format paged attention that keeps recent and protected anchor pages in FP8 while older pages are compressed into packed TQ3, allowing every live‑request page to remain addressable. The approach uses format‑specific kernels and a globally normalized online‑softmax merge to compute partial attention states, enabling direct heterogeneous decoding without a dense shadow cache. Experiments on Qwen3.6‑27B on a 96‑GB NVIDIA RTX PRO 6000 Blackwell GPU show 3.50× compression over BF16 and 1.75× over FP8, with minimal impact on performance across long‑context benchmarks.
The paper introduces LatentPort, a method that allows a language model to transfer its live memory to another model without requiring the receiver to reread the context. Experiments on a Qwen3.5 4B-to-9B sibling pair show that adding a Gated DeltaNet (GDN) persistent-state package reduces negative log‑likelihood by 0.747 nats/token and improves performance across 64 PG19 documents. The study also demonstrates that direct recurrent and convolution reuse outperforms learned GDN maps, and a 434,176‑parameter correction further narrows the performance gap to the native 9B model.
The paper introduces CORE, a method that jointly calibrates coverage and redistributes evicted mass for key–value (KV) cache in long‑context decoding. CORE combines query utility and log‑determinant coverage into a lightweight indexer, using a single calibrated distribution to both retain complementary KV states and parameterize latent‑memory writes. Experiments on three backbones show CORE outperforms the RULER baseline by up to 3.78 points at 90% compression and demonstrates strong decoding efficiency on LongBench and repeated‑eviction tests.