Large language models (LLMs) increasingly need to process long documents and codebases, making long-context attention a major prefill bottleneck. Block sparse attention reduces computation by skipping unimportant blocks of the attention matrix, but scoring blocks accurately can itself be expensive. We propose Delta-Based Tile Scoring (DTS), a training-free, content-adaptive tile scoring method for sparse block attention. DTS selects query and key anchors according to local vector changes, using fewer representatives in redundant regions while retaining more detail in heterogeneous regions. To accelerate anchor selection, parallel delta packing partitions the query and key sequences into tile-aligned chunks and performs bounded local scans over these chunks concurrently. The selected tiles are then computed using the original query, key, and value vectors. On LLaMA-3.1-8B-Instruct, DTS keeps average RULER and LongBench scores within 0.10 percentage points of full attention, with RULER evaluated up to 128K tokens. At 128K tokens, DTS reduces measured scoring-pipeline latency by 38% compared with XAttention. On four NVIDIA H100 GPUs, it achieves a 3.3$\times$ prefill speedup over a PyTorch full-attention baseline using FlashAttention-2 at 512K tokens.
Visual Question Answering (VQA) imposes substantial computational overhead on vision-language models (VLMs), motivating model-agnostic input reduction methods that can be readily applied across diverse VLMs, including in black-box settings. Simply enforcing a visual-token cap reduces computation, but it can discard information needed to answer the question. A single reduction strategy also struggles to accommodate different scenarios. Our key observation is that cropping and compression are complementary: cropping highlights small targets and local details, whereas whole-image compression better preserves context for large targets or distributed evidence. Based on this observation, we propose C$^2$-Routing, which adaptively selects cropping or compression according to the input image and question. Since both strategies may still discard key information, we further design a safety gate that uses the predicted joint correctness distribution under original and reduced inputs to guide fallback and protect accuracy. To accommodate consumer GPU resource constraints, all training and evaluation experiments run on RTX 3090 GPUs. To measure gains beyond token capping, we use full-image baselines with the same cap of 1,024 visual tokens. We conduct extensive evaluation on three VLMs and 18 datasets spanning local details, global scenes, scene text, and documents. With the visual budget already capped, C$^2$-Routing further reduces end-to-end FLOPs by 36.53% and visual tokens by 42.36% on average, with a mean relative task-score decrease of only 0.77%.
Mixture-of-Experts (MoE) scales capacity via conditional computation, but Transformers lack a native knowledge lookup primitive. We introduce conditional memory, instantiated via Deep Sparse Embedding (DSE), which indexes a massive embedding table using local n-grams for retrieval. We formalize sparsity allocation problem—how to split a fixed parameter budget between MoE experts and DSE memory—and find a U-shaped scaling law that identifies an optimal balance. Scaling to 27B parameters, DSE outperform an iso-parameter and iso-FLOPs MoE baseline across knowledge and reasoning benchmarks, and achieve markedly stronger long-context performance. Mechanistic analyses show that DSE offloads early-layer static recall into memory, freeing effective depth and attention for higher-level reasoning. DSE is also infrastructure-efficient: its deterministic hashing enables offloading massive parameters into host memory during inference with negligible throughput overhead.
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表