arXiv:2606. 28560v1 Announce Type: cross Abstract: We study sparse self-attention in which each query attends to a dense local window plus a set of Fibonacci-spaced offsets, with a per-layer scalar alpha that compresses or expands the spacing.
The paper introduces Stratified Sparse Attention (SSA), a framework that blends deterministic top‑k retrieval with score‑guided sampling to reduce the gap between sparse and dense attention in large language models. SSA attends to high‑scoring tokens deterministically while sampling the remaining attention mass using proxy scores, thereby recovering more of the attention distribution. Experiments on benchmarks such as RULER, LongBench, LOFT, and AIME show that SSA outperforms existing sparse methods across models ranging from 3B to 27B parameters, achieving significant accuracy gains and up to 2.96× decode speedup with optimized CUDA kernels.
arXiv:2504. 17768v3 Announce Type: replace-cross Abstract: Sparse attention offers a promising strategy to extend long-context capabilities in Transformer LLMs, yet its efficiency-accuracy trade-offs remain unclear due to the lack of comprehensive evaluation.
arXiv:2606. 13392v1 Announce Type: new Abstract: Ultra-long-context capability is becoming indispensable for frontier LLMs: agentic workflows, repository-scale code reasoning, and persistent memory all require the model to jointly attend over hundreds of thousands to millions of tokens, yet the quadratic cost of softmax attention makes this untenable at deployment scale.
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.
The paper introduces FlashAttention with Statistics, a method that computes attention‑based detector features without materializing the full attention matrix, thereby reducing memory usage and latency. By accumulating statistics during the FlashAttention pass or via sparse lookups, the approach adds only a small overhead—16% at 2K tokens and 1.4% at 35K tokens on an H100 GPU—and stores just 22.5 KiB per response token. Experiments show that frozen Qwen3 proxy models, even as small as 0.6 B parameters, can serve as strong hallucination detectors, outperforming larger zero‑shot judges and specialized grounding checkers on several benchmarks, with attention features further boosting performance.
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.
Large language models (LLMs) exhibit a human‑like working‑memory pattern on the N‑back task, with performance dropping as memory load increases and errors mainly on lure items. Two hypotheses—precision limits of softmax attention and interference from competing items—were tested across 15 pretrained LLMs and minimal transformers. Results show that interference, not precision, constrains LLMs’ working‑memory performance, as suppressing non‑target item identities improves sensitivity and larger models benefit more from reducing interference.
arXiv:2608. 01676v1 Announce Type: cross Abstract: Sparse attention is widely deployed in long-context serving stacks, yet no framework audits how discarding blocks changes the influence of specific content on model output.
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.
arXiv:2607. 13205v1 Announce Type: cross Abstract: Attention-based KV cache eviction (H2O and its descendants) compresses the memory-constrained state of a long-context model by ranking tokens on accumulated attention mass, treated here as signal energy, and keeping the heaviest.
arXiv:2607. 09052v1 Announce Type: new Abstract: Block sparse attention is a hardware friendly way to alleviate the key-value (KV) cache read bottleneck in large language models (LLMs).