CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing

• URL: http://arxiv.org/abs/2609.01925v1
• Date: Tue, 01 Sep 2026 22:49:46 GMT
• Title: CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing
• Abstract summary: We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm.<n>We show that the routing decision can be read directly off the structure of the proxy attention map.<n>We formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts.
• Score: 90.63350360476294
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

Related papers

• Accelerating Unified Multimodal Models with Core-Expansion Routing and Unified Computation Scheduling [61.28599393707943]
  Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps.<n>We propose a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment.<n> Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93times$ end-to-end inference speedup.
  arXiv  Detail & Related papers  (2026-08-29T14:27:55Z)
• The Laws of Context Allocation: Causal Measurement and Closed-Loop Orchestration in Generative Search [19.846305722187328]
  Retrieval-Augmented Generation (RAG) shifts toward diverse portfolio generation.<n>It is stymied by two critical bottlenecks: flawed measurement of evidence utilization, and suboptimal context budget allocation.<n>We introduce a causal leave-one-out probe that accurately isolates generative reliance.<n>We prove that the prevailing strategy of monolithic context widening is an architectural trap penalized by relevance decay.
  arXiv  Detail & Related papers  (2026-08-24T13:44:11Z)
• TRaM-VSR: Importance-Aware Token Routing and Merging for One-Step Diffusion Video Super-Resolution [52.96050253456634]
  TRaM-VSR is a framework for adaptive token allocation using context-aware video priors and network-levels.<n>It is shown that TRaM-VSR accelerates inference significantly while preserving state-of-the-art reconstruction quality and robust temporal consistency.
  arXiv  Detail & Related papers  (2026-07-24T11:57:26Z)
• The Key to Going Linear: Analysis-Driven Transformer Linearization [8.56204304015324]
  We show that softmax relies on key-dependent, rank-1 projections, elucidating why delta-style networks outperform purely gated accumulation.<n>We identify a potential source of approximation errors and introduce structural interventions, specifically sink tokens, short convolutions, and fixed-budget cache routing.
  arXiv  Detail & Related papers  (2026-07-08T17:59:09Z)
• Closed-Form Spectral Regularization for Multi-Task Model Merging [96.82449201305234]
  Model merging combines several independently fine-tuned experts into a single multi-task model without any training data.<n>State-of-the-art merging methods formulate merging as a layer-wise interference problem.<n>We propose SWUDI, a closed-form method that combines a soft exponential filter, which matches the gradient-flow trajectory of iterative descent.
  arXiv  Detail & Related papers  (2026-06-05T14:00:47Z)
• Clipping Bottleneck: Stabilizing RLVR via Stochastic Recovery of Near-Boundary Signals [83.0127582612634]
  Near-boundary Rescue (NSR) is a minimal, plug-and-play modification that retains out-of-bound tokens to recover lost signals.<n>NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO.
  arXiv  Detail & Related papers  (2026-05-21T16:45:31Z)
• AdaFuse: Accelerating Dynamic Adapter Inference via Token-Level Pre-Gating and Fused Kernel Optimization [84.25316984309725]
  Integration of dynamic, sparse structures with parameter-efficient adapters (e.g., LoRA) is a powerful technique for enhancing Large Language Models (LLMs)<n>Despite minimal increases in computational load, the computation latency often skyrockets, leading to decoding speeds slowing by over 2.5 times.<n>AdaFuse is a framework built on a tight co-design between the algorithm and the underlying hardware system to enable efficient dynamic adapter execution.
  arXiv  Detail & Related papers  (2026-03-12T12:46:42Z)
• VSPrefill: Vertical-Slash Sparse Attention with Lightweight Indexing for Long-Context Prefilling [0.0]
  Existing sparse attention methods face a trade-off among context adaptivity, sampling overhead, and fine-tuning costs.<n>We propose VSPrefill, a mechanism requiring lightweight training that uses the vertical-slash structural pattern in attention distributions.<n>VSPrefill preserves 98.35% of the full attention accuracy while delivering a 4.95x average speedup at a context length of 128k.
  arXiv  Detail & Related papers  (2026-03-03T09:24:58Z)
• Adaptive Structured Pruning of Convolutional Neural Networks for Time Series Classification [0.776514389034479]
  We propose Dynamic Structured Pruning (DSP), a fully automatic, structured pruning framework for convolution-based TSC models.<n>Our approach achieves an average compression of 58% for LITETime and 75% for InceptionTime architectures while maintaining classification accuracy.<n>Redundancy analyses confirm that DSP produces compact and informative representations, offering a practical path for scalable and efficient deep TSC deployment.
  arXiv  Detail & Related papers  (2026-02-13T09:18:59Z)
• Long-Context Modeling with Dynamic Hierarchical Sparse Attention for On-Device LLMs [17.499497967319332]
  We introduce Dynamic Hierarchical Sparse Attention (DHSA), a data-driven framework that dynamically predicts attention sparsity online without retraining.<n>DHSA matches dense attention in accuracy, while reducing prefill latency by 20-60% and peak memory usage by 35%.<n>Our experiments on Gemma2 with Needle-in-a-Haystack Test and LongBench show that DHSA matches dense attention in accuracy, while reducing prefill latency by 20-60% and peak memory usage by 35%.
  arXiv  Detail & Related papers  (2025-10-28T16:34:18Z)

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.