The paper introduces On‑Demand Attention (ODA), a local‑first decoding strategy that predicts when a pretrained language model would benefit from global attention. By training only a lightweight recall head, ODA selectively triggers global attention during generation, keeping pretrained weights unchanged and preserving the full key‑value cache for future recall. Experiments on Qwen, Gemma, and hybrid‑attention models show that ODA largely recovers performance lost with local attention while significantly cutting global reads, enabling faster long‑context inference.
The paper investigates how attention dynamics evolve across recurrent depth in language models, finding that attention support stabilizes early while hidden states and outputs take longer. It proposes WISE, a training‑free method that uses full attention in early steps and then reuses the discovered sparse working set for later steps, preserving performance on multi‑hop QA tasks. Experiments show that WISE maintains quality up to 2K context, offers measurable speedups, and highlights the importance of recurrent discovery of attention support.
By Ke Wan, Chen Chen
The paper introduces Declarative Attention (DA), a protocol that lets language models explicitly declare which parts of their context to focus on during generation. By partitioning decoding into full-context, region-specific, and recent-output-only modes, the inference engine can skip large portions of the KV cache, dramatically reducing attended tokens. Experiments on 15 long-context tasks with off-the-shelf models show significant savings (52.0% and 31.1% reductions) with only modest accuracy drops that diminish as model size increases.
By Namgyu Ho, Huzama Ahmad, Woosung Koh, Se-Young Yun, Tal Schuster, Cicero Nogueira dos Santos
arXiv:2603. 17484v2 Announce Type: replace-cross Abstract: Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval.
By Sakshi Choudhary, Aditya Chattopadhyay, Luca Zancato, Elvis Nunez, Matthew Trager, Wei Xia, Stefano Soatto
Elastic Threshold Attention (ETA) is a trainable attention mechanism that dynamically predicts contextual thresholds from query representations, enabling selective pruning of KV cache tokens during long‑context decoding. By multiplicatively suppressing sub‑threshold logits during training, ETA avoids representation collapse and eliminates localized attention sinks, allowing a 1.45B model to match dense attention performance at roughly 85% training sparsity and 38% active decode density. At inference, a custom Triton kernel achieves up to 2.5× faster decoding on sequences up to 512K tokens, and an offline calibration step can further reduce compute by 27% by freezing per‑head thresholds.
By Themistoklis Haris, Henry Li, Maryam Karimzadehgan
arXiv:2606. 09079v1 Announce Type: cross Abstract: Conventional LLMs keep the full KV cache loaded during decoding, causing a severe GPU memory bottleneck for ultra-long context serving.
By Yan Wang, Qifan Zhang, Jiachen Yu, Tian Liang, Dongyang Ma, Xiang Hu, Zibo Lin, Chunyang Li, Zhichao Wang, Jia Li, Yujiu Yang, Haitao Mi, Dong Yu
arXiv:2608. 19920v1 Announce Type: new Abstract: A lot of prior work addressed key-value (KV) cache selection and compression by sparse attention to enable long-context inference for transformer language models without excessive hardware budgets.
By Matthias Seeger, Zeyu Zhang, Vihang Patil, Konstantinos Benidis, Sebastian Schelter
The paper introduces Memory Attention (MA), a new attention mechanism for language models that replaces the traditional value projection with token-indexed memory combined with contextual keys. MA generates values by merging layer‑specific token memory with contextual information, allowing normalization to be folded into memory tables and reducing value construction to a simple lookup and addition. Experiments show that, with matched training token budgets and additional memory parameters, MA improves language modeling performance and average downstream task results across various attention configurations.
By Jiale Kang
arXiv:2608. 05326v1 Announce Type: new Abstract: Autoregressive large language model inference is increasingly constrained by the memory footprint of the Key-Value (KV) cache.
By Ayushman Garg, Akshita Gupta, Shaswata Bhattacharya, Abhishek Gupta, Sandeep Kumar, Manoj Kumar
arXiv:2610.15715v1 Announce Type: new
Abstract: Linear attention promises the efficiency needed for long-context language modelling, yet its fixed-size recurrent state must simultaneously support ong...
By Li Nanbo, Firas Laakom, Yilan Zhang, Piotr Pi\k{e}kos, Yimeng Chen, Liangyu Wang, Yifan Xu, Mateusz Ostaszewski, Dandan Guo, Wenyi Wang, Yanning Dai, Yuhui Wang, J\"{u}rgen Schmidhuber
arXiv:2610.19210v1 Announce Type: new
Abstract: Transformers are efficient to pre-train at scale because attention permits parallel computation across sequence positions, but long-context inference r...
By Sarah Pan, Ali Behrouz
arXiv:2610.20503v1 Announce Type: new
Abstract: Speculative decoding accelerates autoregressive generation by drafting multiple tokens per step, yet feature-based drafters rely on target hidden featu...
By Bohan Sun, Zunhai Su, Xialie Zhuang, Jianjin Li, Ngai Wong, Chuan-Wei Kuo