Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning

• URL: http://arxiv.org/abs/2502.02770v5
• Date: Tue, 04 Nov 2025 12:04:06 GMT
• Title: Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning
• Abstract summary: Twilight is a framework to bring adaptive sparsity to any existing sparse attention algorithm.<n>We show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4times$ acceleration in self-attention operations and $3.9times$ acceleration in end-to-end per token latency.
• Score: 36.25003932215851
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Leveraging attention sparsity to accelerate long-context large language models (LLMs) has been a hot research topic. However, current algorithms such as sparse attention or key-value (KV) cache compression tend to use a fixed budget, which presents a significant challenge during deployment because it fails to account for the dynamic nature of real-world scenarios, where the optimal balance between accuracy and efficiency can vary greatly. In this paper, we find that borrowing top-$p$ sampling (nucleus sampling) to sparse attention can surprisingly achieve adaptive budgeting. Based on this, we propose Twilight, a framework to bring adaptive sparsity to any existing sparse attention algorithm without sacrificing their accuracy. Empirical results show that Twilight can adaptively prune at most 98% of redundant tokens, leading to $15.4\times$ acceleration in self-attention operations and $3.9\times$ acceleration in end-to-end per token latency in long context LLM decoding.

Related papers

• DCP-Prune: Ultra-Low Token Pruning with Distribution Consistency Preservation [55.38884985761757]
  Recent vision token pruning methods effectively preserve model performance under moderate token budgets but become unstable under ultra-low token budget.<n>We propose a two-stage pruning framework consisting of Anchor-Context Graph Recovery (ACGR) and Text-Aware Token Cluster Selection (TATCS)<n>Experiments demonstrate that our method achieves superior and more stable performance under ultra-low token budget.
  arXiv  Detail & Related papers  (2026-06-15T12:24:07Z)
• Agile Online Model Selection: Resolving Adaptation Lag via Safeguarded Large Learning Rates [9.463718838347955]
  We propose a novel optimistic online mirror descent that utilizes safeguarded large learning rates up to $(T)$.<n>Our key technical contribution is a post-hoc penalty mechanism that dynamically monitors unstable updates and excludes learning rates incurring excessive regret.<n> Empirical evaluations on three synthetic and eleven diverse real-world datasets demonstrate that our approach reduces the adaptation lag from hundreds of rounds to a few rounds.
  arXiv  Detail & Related papers  (2026-05-26T12:18:08Z)
• Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps [12.944531570933854]
  Long-context inference in large language models is bottlenecked by the quadratic cost of full attention.<n>We show that full-attention LLMs are already intrinsically sparse and can be transformed into highly sparse models with only minimal adaptation.<n>By exploiting the model's intrinsic sparsity, RTPurbo achieves sparsification with only a few hundred training steps.
  arXiv  Detail & Related papers  (2026-05-16T10:51:58Z)
• Scaling Attention via Feature Sparsity [50.64995497733461]
  Scaling Transformers to ultra-long contexts is bottlenecked by the $O(n2 d)$ cost of self-attention.<n>We propose Sparse Feature Attention, where queries and keys are represented as $k$-sparse codes that preserve high-dimensional expressivity.<n>Across GPT-2 and Qwen3 pretraining, SFA matches dense baselines while improving speed by up to $2.5times$ and reducing FLOPs and KV-cache by nearly 50%.
  arXiv  Detail & Related papers  (2026-03-17T08:41:50Z)
• $\ abla$-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space [71.23672814629448]
  $nabla$-Reasoner is an iterative generation framework that integrates differentiable optimization over token logits into the decoding loop.<n>$nabla$-Reasoner achieves over 20% accuracy improvement on a challenging mathematical reasoning benchmark.
  arXiv  Detail & Related papers  (2026-03-05T08:42:54Z)
• Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction [19.14455067106419]
  Current KV cache eviction methods rely on instantaneous metrics, implicitly assuming that score magnitudes are consistent proxies for importance across all heads.<n>In this paper, we propose that optimal budget allocation should be governed by the marginal utility in preserving long-term semantic information.<n>We implement a data-driven offline profiling protocol to facilitate the practical deployment of LU-KV.
  arXiv  Detail & Related papers  (2026-02-09T12:23:38Z)
• Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks [87.16809558673403]
  Spiking neural networks (SNNs) compute with discrete spikes and exploit temporal structure.<n>We study a timing-only adversary that retimes existing spikes while preserving spike counts and amplitudes in event-driven SNNs.
  arXiv  Detail & Related papers  (2026-02-03T09:06:53Z)
• Predictive Scheduling for Efficient Inference-Time Reasoning in Large Language Models [6.002670452103349]
  Large language models (LLMs) achieve state-of-the-art accuracy on complex reasoning tasks.<n>But using a fixed token budget per query leads to over-computation on easy inputs and under-computation on hard ones.<n>We introduce Predictive Scheduling, a plug-and-play framework that pre-runs lightweight predictors to estimate each query's optimal reasoning length or difficulty before any full generation.
  arXiv  Detail & Related papers  (2026-02-01T13:58:23Z)
• vAttention: Verified Sparse Attention [100.98210818821688]
  vAttention is a practical sparse attention mechanism with user-specified $(epsilon, delta)$ guarantees on approximation accuracy (thus, verified)<n>We show that vAttention significantly improves the quality of sparse attention across datasets.<n>It can be deployed in reasoning scenarios to achieve fast decoding without compromising model quality.
  arXiv  Detail & Related papers  (2025-10-07T08:46:08Z)
• R-Stitch: Dynamic Trajectory Stitching for Efficient Reasoning [80.104336426172]
  Chain-of-thought (CoT) enhances problem-solving ability of large language models.<n>CoT incurs substantial inference cost due to long autoregressive trajectories.<n>We introduce R-Stitch, a training-free hybrid decoding framework.
  arXiv  Detail & Related papers  (2025-07-23T08:14:36Z)
• Steering LLM Thinking with Budget Guidance [48.65894557568655]
  Budget guidance is a method for steering the reasoning process of LLMs toward a target budget without requiring any fine-tuning.<n>Our approach introduces a lightweight predictor that models a Gamma distribution over the remaining thinking length.<n>This signal is then used to guide generation in a soft, token-level manner, ensuring that the overall reasoning trace adheres to the specified thinking budget.
  arXiv  Detail & Related papers  (2025-06-16T17:57:05Z)
• $\ exttt{SPECS}$: Faster Test-Time Scaling through Speculative Drafts [55.231201692232894]
  $textttSPECS$ is a latency-aware test-time scaling method inspired by speculative decoding.<n>Our results show that $textttSPECS$matches or surpasses beam search accuracy while reducing latency by up to $sim$19.1%.
  arXiv  Detail & Related papers  (2025-06-15T05:50:05Z)
• Delta Attention: Fast and Accurate Sparse Attention Inference by Delta Correction [52.14200610448542]
  A transformer has a quadratic complexity, leading to high inference costs and latency for long sequences.<n>We propose a simple, novel, and effective procedure for correcting this distributional shift.<n>Our method can maintain approximately 98.5% sparsity over full quadratic attention, making our model 32 times faster than Flash Attention 2 when processing 1M token prefills.
  arXiv  Detail & Related papers  (2025-05-16T13:48:33Z)
• Sparsity Forcing: Reinforcing Token Sparsity of MLLMs [40.93786579652003]
  We explicitly reinforce token sparsity in well-posed multimodal large language models (MLLMs) through a simple RL-based post-training framework named textitSparsity Forcing.<n>Our method explores the efficiency-accuracy trade-off by running multiple rollouts with different token budgets, where both efficiency (token reduction ratio) and performance (answer correctness) are formulated as joint rewards.
  arXiv  Detail & Related papers  (2025-04-23T01:45:55Z)
• $φ$-Decoding: Adaptive Foresight Sampling for Balanced Inference-Time Exploration and Exploitation [22.607133083903125]
  In-time optimization scales computation to derive deliberate reasoning steps for effective performance. We frame the decoding strategy as foresight sampling, leveraging simulated future steps to obtain globally optimal step estimation. Experiments show $phi$-Decoding outperforms strong baselines in both performance and efficiency.
  arXiv  Detail & Related papers  (2025-03-17T15:38:33Z)
• PowerAttention: Exponentially Scaling of Receptive Fields for Effective Sparse Attention [73.26995918610669]
  Large Language Models (LLMs) face efficiency bottlenecks due to the quadratic complexity of the attention mechanism when processing long contexts. We introduce PowerAttention, a novel sparse attention design that facilitates effective and complete context extension. Experiments demonstrate that PowerAttention outperforms existing static sparse attention methods by $5sim 40%$.
  arXiv  Detail & Related papers  (2025-03-05T15:24:11Z)
• Progressive Sparse Attention: Algorithm and System Co-design for Efficient Attention in LLM Serving [10.835583587146274]
  This paper presents PSA, a $underlineP$rogressive $underlineS$parse $underlineA$ttention mechanism. It integrates algorithmic innovations with system co-design to achieve both high inference accuracy and improved efficiency in large language models. Experiments demonstrate that PSA reduces KV cache usage for attention computation by up to 2.4$times$ and 8.8$times$, and increases end-to-end serving throughput by up to 1.4$times$ and 2.0$times$.
  arXiv  Detail & Related papers  (2025-03-01T07:56:42Z)
• Activation-aware Probe-Query: Effective Key-Value Retrieval for Long-Context LLMs Inference [56.71209737306054]
  We propose textbfActQKV, a training-free, textbfActivation-aware approach that dynamically determines probe-textbfQuery and leverages it to retrieve the relevant textbfKV pairs for inference. Experiments on the Long-Bench and $infty$ Benchmarks demonstrate its state-of-the-art performance with competitive inference quality and resource efficiency.
  arXiv  Detail & Related papers  (2025-02-19T08:50:44Z)
• Tactic: Adaptive Sparse Attention with Clustering and Distribution Fitting for Long-Context LLMs [10.52833484759311]
  We propose Tactic, a sparsity-adaptive and calibration-free sparse attention mechanism. It dynamically selects tokens based on their cumulative attention scores rather than a fixed token budget. We show that Tactic outperforms existing sparse attention algorithms, achieving superior accuracy and up to 7.29x decode attention speedup.
  arXiv  Detail & Related papers  (2025-02-17T08:39:43Z)
• AttentionPredictor: Temporal Pattern Matters for Efficient LLM Inference [51.1972443343829]
  We propose AttentionPredictor, which is the first learning-based critical token identification approach. AttentionPredictor accurately predicts the attention score while consuming negligible memory. We also propose a cross-token critical cache prefetching framework that hides the token time overhead to accelerate the decoding stage.
  arXiv  Detail & Related papers  (2025-02-06T13:41:46Z)
• Squeezed Attention: Accelerating Long Context Length LLM Inference [61.787865959140994]
  We propose Squeezed Attention to accelerate applications where a large portion of the input context is fixed.<n>During inference, we compare query tokens from the user input with the centroids to predict which keys from the fixed context are semantically relevant.<n>We also present a hierarchical version of our algorithm which can reduce the complexity of attention from linear to logarithmic with respect to the fixed context length.
  arXiv  Detail & Related papers  (2024-11-14T18:54:19Z)
• ZipVL: Efficient Large Vision-Language Models with Dynamic Token Sparsification [29.163757099307553]
  The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase.<n>We present ZipVL, an efficient inference framework designed for LVLMs through a dynamic ratio allocation strategy of important tokens.
  arXiv  Detail & Related papers  (2024-10-11T07:24:21Z)
• Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inference [19.447729423696096]
  Large Language Models have excelled in various domains but face efficiency challenges due to the growing Key-Value (KV) cache.<n>Recent efforts aim to reduce KV cache size by evicting vast non-critical cache elements during runtime.<n>We propose Ada-KV, the first head-wise adaptive budget allocation strategy.
  arXiv  Detail & Related papers  (2024-07-16T09:53:32Z)
• Sparser is Faster and Less is More: Efficient Sparse Attention for Long-Range Transformers [58.5711048151424]
  We introduce SPARSEK Attention, a novel sparse attention mechanism designed to overcome computational and memory obstacles. Our approach integrates a scoring network and a differentiable top-k mask operator, SPARSEK, to select a constant number of KV pairs for each query. Experimental results reveal that SPARSEK Attention outperforms previous sparse attention methods.
  arXiv  Detail & Related papers  (2024-06-24T15:55:59Z)
• SampleAttention: Near-Lossless Acceleration of Long Context LLM Inference with Adaptive Structured Sparse Attention [47.5772915135952]
  Large language models (LLMs) now support extremely long context windows. The quadratic complexity of vanilla attention results in significantly long Time-to-First-Token (TTFT) latency. We propose SampleAttention, an adaptive structured and near-lossless sparse attention.
  arXiv  Detail & Related papers  (2024-06-17T11:05:15Z)
• Adaptive importance sampling for heavy-tailed distributions via $\alpha$-divergence minimization [2.879807093604632]
  We propose an AIS algorithm that approximates the target by Student-t proposal distributions. We adapt location and scale parameters by matching the escort moments of the target and the proposal. These updates minimize the $alpha$-divergence between the target and the proposal, thereby connecting with variational inference.
  arXiv  Detail & Related papers  (2023-10-25T14:07:08Z)
• Latency-aware adaptive shot allocation for run-time efficient variational quantum algorithms [0.0]
  adaptive strategies have been proposed to determine the number of gradient measurement shots used to estimate the iteration. We present an adaptive strategy that the number of shots in each balance to maximize expected gain per unit time or cost by explicitly taking into account the overhead. Numerical simulations show that our adaptive shot strategy is actually efficient for Adam, outperforming many existing state-of-the-art adaptive shot strategies.
  arXiv  Detail & Related papers  (2023-02-09T03:27:25Z)
• AdapLeR: Speeding up Inference by Adaptive Length Reduction [15.57872065467772]
  We propose a novel approach for reducing the computational cost of BERT with minimal loss in downstream performance. Our method dynamically eliminates less contributing tokens through layers, resulting in shorter lengths and consequently lower computational cost. Our experiments on several diverse classification tasks show speedups up to 22x during inference time without much sacrifice in performance.
  arXiv  Detail & Related papers  (2022-03-16T23:41:38Z)
• Large-scale Optimization of Partial AUC in a Range of False Positive Rates [51.12047280149546]
  The area under the ROC curve (AUC) is one of the most widely used performance measures for classification models in machine learning. We develop an efficient approximated gradient descent method based on recent practical envelope smoothing technique. Our proposed algorithm can also be used to minimize the sum of some ranked range loss, which also lacks efficient solvers.
  arXiv  Detail & Related papers  (2022-03-03T03:46:18Z)
• Momentum Accelerates the Convergence of Stochastic AUPRC Maximization [80.8226518642952]
  We study optimization of areas under precision-recall curves (AUPRC), which is widely used for imbalanced tasks. We develop novel momentum methods with a better iteration of $O (1/epsilon4)$ for finding an $epsilon$stationary solution. We also design a novel family of adaptive methods with the same complexity of $O (1/epsilon4)$, which enjoy faster convergence in practice.
  arXiv  Detail & Related papers  (2021-07-02T16:21:52Z)
• Layer-adaptive sparsity for the Magnitude-based Pruning [88.37510230946478]
  We propose a novel importance score for global pruning, coined layer-adaptive magnitude-based pruning (LAMP) score. LAMP consistently outperforms popular existing schemes for layerwise sparsity selection.
  arXiv  Detail & Related papers  (2020-10-15T09:14:02Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.