2026· Annual Meeting of the Association for Computational Linguistics· pp. 1028-1044· 0 citations· 34 references
Computer Science
TL;DR
This paper introduces RoPE-based Block-wise Sparse Attention (RoBSA), a method designed specifically for MLA during the decoding stage of model inference that significantly reduces end-to-end inference latency in the decoding stage by up to 2 .
Abstract
Large Language Models (LLMs) have rapidly advanced in recent years, scaling up in both parameter count and context length. However, as context windows extend from thousands to hundreds of thousands of tokens, attention computation becomes the dominant source of memory usage and runtime in decoding stages, severely limiting the efficiency and scalability of long-context LLMs. Sparse attention has emerged as a promising solution, reducing complexity by computing attention over only a subset of context tokens. However, the sparse attention for Multi-head Latent Attention (MLA) which is a variant of standard MHA is rarely studied. In this paper, we introduce RoPE-based Block-wise Sparse Attention (RoBSA), a method designed specifically for MLA during the decoding stage of model inference. RoBSA leverages the decoupled nature of RoPE within MLA to implement token selection in a blockwise manner. RoBSA is a lightweight, training-free, and layer-aware algorithm that can be integrated in a plug-and-play fashion. Our method significantly reduces end-to-end inference latency in the decoding stage by up to 2 . 55 × with minimal accuracy loss compared to full attention in long-context scenarios for very large models.
Scaling modern large language models (LLMs) to long contexts is limited by the quadratic computation cost, and poor length extrapolation of dense attention. Chunk-wise sparse attention offers a promising alternative, but all existing methods fall short of full attention because of their inaccurate chunk selection. We propose Hierarchical Landmark Sparse (HiLS) Attention, a chunk-wise sparse attention mechanism that learns chunk selection end-to-end under the language-modeling (LM) loss. HiLS factorizes attention hierarchically: each query performs attention independently with each retrieved chunk to extract chunk-specific information, and the resulting outputs are fused according to chunk retrieval scores. By incorporating retrieval scores into the forward attention computation, HiLS optimizes them directly with the LM loss, enabling end-to-end retrieval learning and native sparse training. Experimental results show that HiLS-Attention achieves performance comparable to, and in some cases better than, full attention at in-domain context lengths. Meanwhile, HiLS-Attention extrapolates more than $64\times$ the training context length with 90% retrieval accuracy, far beyond full attention. Moreover, existing full-attention models can be converted to HiLS-Attention with lightweight continued pretraining, preserving in-domain performance while acquiring ultra-long-context extrapolation. Together with its sparse KV access and computation, HiLS-Attention breaks the usual efficiency-performance trade-off, enabling long-context LLMs that are both more efficient and more effective on general long-context tasks than their full-attention counterparts.
Block-sparse attention scales long-context language models by replacing the O(N^2) softmax with a per-query top-k selection over key blocks. This cutoff is myopic: when the k-th and (k+1)-th blocks are nearly tied in score, the selector commits without spending extra budget, and a dropped block carrying answer evidence is unrecoverable downstream. We propose a value-of-information router that measures, for each query, how decisively the top-k cut was made, and doubles the kept set for the queries where that gap is smallest; the rule is backbone-agnostic and stacks with existing block-scoring methods such as Quest. On LongBench-v2 medium at n=215 (the entire dataset subset), router-on-Quest reaches paired recall 0.75 vs. top-k 0.47 -- +28 pp over the SSA-style baseline (McNemar p<0.01) -- and lands within 2 pp of dense on RULER NIAH multikey at the same context. The lift reproduces on four models from three architectures (Qwen2.5, Mistral-Nemo, Qwen3.6). At 128K, the router preserves 0.81 and 0.89 of dense accuracy on Qwen2.5-7B-1M and Qwen3.6 (vs. SSA-style top-k at 0.09 on the former) while the fused selection-plus-kernel pipeline runs at 0.62x and 0.80x dense wall time.
Pruning is essential for the efficient deployment of Large Language Models (LLMs); however, it causes severe performance degradation due to the structural distortion induced by sparsity. Existing recovery strategies, such as LoRA, predominantly employ global fine-tuning, often overlooking the mechanistic root of this degradation: the layer-wise accumulation and amplification of local errors. To address this limitation, we propose LaCo ( La yer-wise Co mpensation), a framework that reori-ents the recovery paradigm from global adaptation to hierarchical representation alignment. By sequentially optimizing each layer to re-construct the model’s hidden states, LaCo effectively intercepts the error propagation chain at its source. Extensive experiments demonstrate that LaCo surpasses parameter-efficient baselines in both perplexity reduction and zero-shot reasoning. Notably, it reduces recovery-time memory usage to approximately 1 / 7 of the baseline and requires only 2,048 unlabeled samples to match a LoRA model trained on 50k examples—achieving a ∼ 25 × improvement in data efficiency.
Yingen Liu, Fan Wu, Xuyan Pan et al.· Annual Meeting of the Associ...· 0 citations
Multi-head Latent Attention (MLA), introduced in DeepSeek-V2, compresses key-value pairs through a shared low-rank bottleneck (cKV), achieving 81% KV-cache reduction during inference. Despite its adoption in massive production models, no prior work has studied what information this bottleneck preserves or discards, nor how it reshapes internal transformer circuits. We present the first comprehensive mechanistic interpretability study of MLA, training a 114M-parameter transformer (pretrained on a web/code/math mixture, fine-tuned on TinyStories) and analyzing its representations through SVD, attention head taxonomy, linear probing, and a disruption-attribution analysis. Our key findings are: (1) the cKV bottleneck learns a pure content representation, preserving entity identity (98% retention) while discarding positional information, validating MLA's separation of content from position via RoPE; (2) induction heads co-locate at a single layer (Layer 12), unlike their distributed formation in standard MHA; (3) a single"semantic hub"layer (Layer 15) simultaneously exhibits the highest SVD effective rank and strongest disruption-attribution score; and (4) the bottleneck is globally over-provisioned, using only 46% of its capacity on average. These findings suggest MLA does not merely compress attention passively, but reshapes how the model organizes content, position, and circuit structure. We view this as an initial data point and detail scope limitations in Section 5.
S. Dhruvil, Fenil Sojitra, Ravirajsinh Chauhan· 0 citations
The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation. Code is available at https://github.com/Tencent/AngelSlim.
Learning effectively from limited data is critical in domains like security where labeled examples are scarce. Large language models (LLMs) have demonstrated some capabilities for data-efficient learning, especially through parameter-efficient adaptation methods, but continue to struggle when faced with few samples for difficult tasks. To meet this challenge, we propose Attention Head Reweighting (AHR), a data-efficient method that adapts LLMs to new text-classification tasks by learning only a single scalar per attention head. This drastically reduces the number of parameters that need to be learned by making use of the functional specialization of individual attention heads. Experiments on diverse open-source text classification datasets show that AHR can outperform standard baselines like LoRA when learning from limited samples, despite having 200-1000x fewer trainable parameters, as our AHR only modifies ~0.0001% of the model's parameters. In addition, our learned weights are easy to interpret and can be analyzed to better understand the mechanisms and attention heads responsible for in-context learning abilities in LLMs.
Tuomas P. Oikarinen, Zixiao Chen, Charlotte Siska et al.· 0 citations