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:2609.08189v1 Announce Type: new
Abstract: Dynamic layer routing reduces the inference cost of Large Language Models (LLMs) by learning to skip layers for individual tokens. Existing methods, ho...
By Hongjin Lin, Wentao Wan, Keze Wang
Attention-Aware Routing (AAR) augments the router in Mixture-of-Experts language models with temporal and spectral features derived from a sliding window of attention weights, thereby separating contextual information from the token’s hidden state. By keeping the base transformer frozen and training only routing parameters, AAR achieves a +3.37‑point improvement on GSM8K over a routing‑only baseline and demonstrates that routing changes propagate through the residual stream to reshape attention without directly updating the attention mechanism. The method also reduces long diverging generations, shows depth‑sensitivity affecting retrieval versus reasoning, and offers a controlled probe of routing‑relevant information across layers.
By Despoina Kosmopoulou, Anastasios Tsetsilas, Efthymios Georgiou, Giannis Karamanolakis, Swastik Roy, Alexandros Potamianos
arXiv:2606. 04511v1 Announce Type: cross Abstract: Sparse attention reduces compute and memory bandwidth for long-context LLM inference.
By Yaosheng Fu, Guangxuan Xiao, Xin Dong, Song Han, Oreste Villa
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:2608. 19758v1 Announce Type: new Abstract: Long-context modeling is a pivotal capability for Large Language Models, yet the quadratic complexity of attention remains a critical bottleneck, particularly during the compute-intensive prefilling phase.
By Qihang Fan, Huaibo Huang, Zhiying Wu, Bingning Wang, Ran He
arXiv:2606. 09514v1 Announce Type: new Abstract: Large language models (LLMs) incur high inference cost due to their depth and parameter scale.
By Yuhua Zhou, Shaoqi Yu, Shichao Weng, Changhai Zhou, Mingze Yin, Fei Yang, Aimin Pan
MoEZip is a routing‑aware key‑value (KV) cache compression technique designed for sparse Mixture‑of‑Experts (MoE) language models. It introduces a Fisher‑information‑based metric to assess how sensitive expert routing is to cache perturbations, and combines this with attention weights to score contexts for eviction. Experiments demonstrate that, even with only 10% of the KV cache retained, MoEZip outperforms state‑of‑the‑art baselines by significant margins on LongBench and RULER benchmarks.
By Minsang Kim, Seungjun Baek
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
arXiv:2606. 06467v1 Announce Type: cross Abstract: Long-context inference in modern LLMs is increasingly constrained by decoding efficiency, especially in reasoning-heavy settings where models generate long intermediate chains of thought.
By Yutao Sun, Yanqi Zhang, Li Dong, Jianyong Wang, Furu Wei
arXiv:2609.22884v1 Announce Type: new
Abstract: Long-context inference has become a defining capability of large language models, but exact dense attention remains costly due to its quadratic scaling...
By Xinwei Long, Weigao Sun, Weibo Gao, Pengkun Jiao, Biqing Qi, Feida Zhu, Yiran Zhong, Steven Hoi, Bowen Zhou
Prefilling-dLLM v2 is a diffusion language model that learns to route computation using a native mask token, enabling efficient sparse attention across long contexts. By appending a mask at the end of each chunk and using a learned role offset, the model selectively attends to top‑K chunks every fourth layer while maintaining local windows elsewhere. Evaluated on 16K–128K contexts, Prefilling-dLLM v2 outperforms other sparse methods (NSA, DSA, SWA) on RULER tasks and achieves a 172× speedup over full dense attention at 128K tokens.