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.
arXiv:2606. 26472v1 Announce Type: new Abstract: As reasoning models emit chains of thought tens of thousands of tokens long, KV cache increasingly becomes a deployment bottleneck.
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.
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 Speculative Probing, a method that repurposes the speculative‑decoding module of large language models for real‑time classification tasks. By appending a trained soft prompt to the target sequence, the approach leverages the already‑cached KV store during inference, adding negligible overhead while achieving higher accuracy than traditional hidden‑state probes. Experiments on four classification tasks across multiple models show that these lightweight probes outperform zero‑shot GPT‑5.4‑mini and rival or surpass specialized 8B safety classifiers without running a full LLM.
RBS-Attention introduces a training‑free, radius‑bounded sparse prefill strategy for long‑context large language models, addressing the mean dilution problem where a block centroid can miss highly relevant tokens. The method employs two complementary selection branches: a centroid base branch that captures average relevance and a rescue branch that uses the maximum key‑block radius to flag under‑estimated blocks. Experiments on Qwen3 models demonstrate significant speedups—over 20× in standalone prefill‑attention and nearly 6× in end‑to‑first‑token time—while maintaining competitive accuracy compared to dense attention.
The paper introduces OpenAttention, a sparse attention framework that uses a single count‑aware mean proxy to represent the unread tail of the key‑value cache, outperforming sampling methods at low read rates. It combines clustering, an ErrorGate that allocates reads based on estimated weight error, and a query‑geometry selector to efficiently compress keys while preserving query direction. Experiments on RULER‑32K HARD with Llama‑3.1‑8B show OpenAttention retains 99 % of dense accuracy at a 1 % read and achieves state‑of‑the‑art accuracy across models, contexts, and datasets, with a custom kernel that is ten times faster than dense attention at the same read rate.
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:2610.14754v1 Announce Type: new Abstract: Hybrid language models interleave Gated DeltaNet (GDN) layers, whose recurrent memory has a fixed size, with full-attention layers, whose key-value (KV...
arXiv:2603. 06642v2 Announce Type: replace-cross Abstract: Test-Time Training (TTT) language models replace the KV-cache with fast weights updated during inference, achieving O(1) memory but suffering catastrophic failure on exact-recall tasks.
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.
arXiv:2602.06317v3 Announce Type: replace Abstract: Attention-mass concentration creates an opportunity for sparse decoding, but retained mass alone does not guarantee a stable greedy decision: retri...
arXiv:2607. 21692v1 Announce Type: new Abstract: Sparse attention reduces the cost of long contexts by allowing each query to read only selected parts of the input.