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
CompKV introduces a compensation‑aware sparse attention framework for long‑context LLM inference. It partitions tokens into blocks and optimizes token selection to minimize the error introduced by block‑level mean compensation, using compact block‑level statistics. Experiments on RULER and LongBench‑Pro show CompKV outperforms other sparse baselines and achieves up to a 6.85× speedup over full attention.
By Zhen Huang, Ruizhe Yao, Danyi Liu, Xinrui Chen, Shuwei Li, Siru Zhong, Zijian Cao, Yushan Lai, Mingming Guo, Weijie Zheng, Haohuan Fu
Large language models (LLMs) often process entire input contexts even when only a small portion is needed to answer a query. The authors introduce dynamic context cutoff, a technique that lets LLMs self‑terminate processing once they have gathered enough task‑relevant information. By analyzing model internals, they find that certain attention heads encode latent sufficiency signals that can be detected with lightweight classifiers, enabling efficient context handling. Experiments on six QA datasets with LLaMA, Qwen, and Mistral models (1B–70B) show a 3.4 % accuracy gain and a 1.33× reduction in token usage, outperforming other context‑efficiency methods at similar token reductions. The study also notes that larger models can intrinsically assess sufficiency through prompting, while smaller models need explicit probing.
By Roy Xie, Junlin Wang, Paul Rosu, Chunyuan Deng, Bolun Sun, Zihao Lin, Bhuwan Dhingra
arXiv:2606. 18587v1 Announce Type: cross Abstract: Decoder-only Transformers compute attention over the KV cache of preceding tokens.
By Zhiyuan Wang, Xuan Luo, Sirui Zeng, Xifeng Yan
The paper introduces On‑Demand Attention (ODA), a decoding strategy that lets pretrained language models decide when to use global attention based on a lightweight recall head. ODA keeps the original model weights unchanged, only training the recall head, and can be implemented with GPU‑side conditional execution to reduce global reads. Experiments on Qwen, Gemma, and hybrid‑attention models show that ODA largely recovers performance lost by local attention while cutting the number of global attention operations.
By Haibo Feng, Ruiqi Liang, Hanyang Peng, Shiqi Yu
arXiv:2506.19549v2 Announce Type: replace-cross
Abstract: Identifying which tokens and activations truly influence model predictions is critical for both the efficiency and interpretability of large...
By Debabrata Mahapatra, Shubham Agarwal, Apoorv Saxena, Subrata Mitra
The paper introduces Pre-hoc Sparsity (PrHS), a method that selects key-value (KV) cache entries before attention scoring to avoid posterior bias in large language model inference. By bounding mutual‑information loss through the dropped attention mass, PrHS offers explicit accuracy control and implements three orthogonal selectors across time, depth, and layer. Experiments on LLaMA and Mistral models show that PrHS cuts retrieval overhead by over 90%, achieves higher sparsity than HShare, and delivers significant speedups and reduced FLOPs on NVIDIA A100 GPUs while maintaining near‑dense accuracy.
By Yifei Gao, Lei Wang, Rong-Cheng Tu, Qixin Zhang, Jun Cheng, Dacheng Tao
The paper investigates how post‑softmax attention weights behave across different language models by marginalizing over query positions to create a joint token‑head space. Across more than 60 diverse LLMs, token‑wise marginal attention reveals a text‑intrinsic signal that is conserved across models, while head‑wise marginal attention yields a model‑private signature that is consistent across documents. This insight enables a practical method for estimating per‑head budgets for key‑value cache eviction, achieving competitive performance on standard benchmarks without per‑document recomputation or training.
By Valentino Maiorca, Walter Nelson, Francesco Locatello
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
arXiv:2607. 02980v1 Announce Type: cross Abstract: Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention.
By Xiang Hu, Xinyu Wei, Hao Gu, Minshen Zhang, Tian Liang, Huayang Li, Lei Zhu, Yan Wang, Sirui Han, Yushi Bai, Kewei Tu, Haitao Mi, Leo Liang
arXiv:2604. 20920v2 Announce Type: replace Abstract: Sparse attention can reduce the cost of long-context inference, but most variants introduce new architectural components.
By Yuzhen Mao, Michael Y. Li, Emily B. Fox
The paper investigates whether recent attention‑mechanism improvements—specifically gated attention, Kimi K3, Kimi Delta Attention, and Attention Residuals—effectively eliminate the attention‑sink problem when scaling language models to a one‑million‑token context window. Using a new diagnostic suite called SinkProbe, the authors evaluate sink mass, massive activation, position‑resolved recall, and the recency gap across four small models that vary only in token mixing and depth. Their findings show that the training objective, rather than the architecture, drives the emergence of attention sinks; gating did not replicate its previously reported benefits at the larger scale, and sink mass, activations, and positional bias behaved independently.
By Sara Rizwan, Samaanah Abdus Salam