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 AMOR, an Adaptive Metacognitive Output Router that selectively activates attention in recurrent-attention hybrid models based on predictive uncertainty. By gating attention blocks with an entropy threshold derived from a running batch median and scaled standard deviation, AMOR reduces attention usage to about 40% of positions while achieving superior common-sense reasoning performance across multiple scales. The approach also improves retrieval accuracy over pure recurrent models and maintains long-context robustness, outperforming fixed-schedule hybrids under distribution shift.
By Haoran Zheng, Chen Shani
The paper reports that chain‑of‑thought (CoT) supervised fine‑tuning (SFT) improves reasoning but systematically harms long‑context recall in hybrid linear‑attention models such as HypeNet and Jet‑Nemotron. Retrieval performance on the Needle‑In‑A‑Haystack benchmark drops dramatically after CoT‑SFT, especially with harder settings and longer contexts. The authors introduce QK‑Restore, a training‑free method that reinstates the query‑key projection matrices from the pre‑SFT checkpoint, which recovers long‑range recall while preserving reasoning gains.
By Xinyu Zhou, Boyu Zhu, Yi Xu, Zhiwei Li, Yingfa Chen, Huiming Wang, Zhijiang Guo
The paper introduces Post-Boundary Bridge (PBB), a hybrid transformer architecture that keeps causal attention within blocks while adding direct connections across block boundaries. PBB focuses on within-block modeling and nearby exchange, delegating long-range communication to full-attention layers. Experiments on dense and mixture-of-experts models ranging from 205 million to 2.07 billion parameters show that PBB hybrids maintain near-full perplexity, competitive downstream performance, and improved source retrieval, while Flash-PBB achieves 1.82× faster decoding with half the local key‑value cache compared to Flash‑SWA.
By Zhibo Yang
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:2607. 06601v1 Announce Type: cross Abstract: Conditional computation can decouple language model quality from per-token inference cost, yet leading techniques act on a single axis in isolation: Mixture-of-Experts (MoE) sparsifies the FFN, Mixture-of-Depths (MoD) skips whole transformer blocks, and KV-cache quantization compresses attention memory.
By Andrii Balashov, Olena Ponomarova
arXiv:2609.15131v1 Announce Type: cross
Abstract: Multimodal large language models (MLLMs) require substantial computation to process numerous visual tokens across all transformer layers. Most method...
By Yuyao Sun, Tao Deng, Shuang Li, Deqing Wang
CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling) is a new method for long-context LLM inference that replaces costly quadratic attention prefilling with a dynamic, input-adaptive sparse routing scheme. It introduces a structural proxy, C_struct, to directly read routing decisions from the proxy attention map, eliminating the need for pooled matrix multiplication and KL divergence. Additionally, CRISP addresses the post-softmax mass cliff by using a sink-aware threshold based on the noise floor, theoretically reducing background noise accumulation to O(n). Empirical results on InfiniteBench, RULER, and LongBench show that CRISP outperforms existing sparse methods and can match or exceed exact dense attention, achieving up to a 5.30× speedup at 512k tokens and significant gains on retrieval-heavy tasks.
By Huu Huy Nguyen, Chien Van Nguyen, Franck Dernoncourt, Ryan A. Rossi, Linh Ngo Van, Jieyang Chen, Thien Huu Nguyen
arXiv:2607. 18363v1 Announce Type: cross Abstract: Feed-forward networks hold two thirds of a transformer's non-embedding parameters, yet the architecture has not received a necessity test that controls parameters, compute, and depth at once.
By Henry Ndubuaku, Karen Mosoyan, Jakub Mroz, Noah Cylich, Satyajit Kumar, Parkirat Sandhu, Roman Shemet, Justin H Lee
The paper introduces OASIS, a method designed to stabilize dual‑normalized attention‑residual architectures by employing explicit null routing and token‑to‑depth null coupling. OASIS mitigates attention sinks and activation outliers, improving low‑bit quantization performance across several language‑model backbones. Empirical results show significant reductions in attention norms and perplexity, with notable gains on long‑context benchmarks.
By Haozheng Luo, Haoran Dai, Jingyuan Huang, Shaoyang Zhang, Xi Chen, Eric Hanchen Jiang, Yijiang Li, Chenghao Qiu, Chenwei Xu, Zhenyu Pan, Haotian Zhang, Binghui Wang, Yan Chen
arXiv:2609.13141v1 Announce Type: new
Abstract: Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context u...
By Zhiwei Li, Lei Zhu, Hao Gu, Xiang Hu, Yan Wang, Haitao Mi, Sirui Han, Leo Liang, Zhijiang Guo
arXiv:2604.23036v2 Announce Type: replace-cross
Abstract: Despite MoE models leading many benchmarks, supervised fine-tuning (SFT) for the MoE architectures remains difficult because its router layer...
By Haoze He, Xingyuan Ding, Xuan Jiang, Xinkai Zou, Alex Cheng, Yibo Zhao, Juncheng Billy Li, Heather Miller