arXiv subjects

Yushan Lai

Publications and source records attributed to Yushan Lai.

3 recordsLinked to original sources

arXiv2026

CompKV: Compensation-Aware KV Selection for Long-Context LLM Inference

Despite their strong performance, large language models (LLMs) are bottlenecked by KV cache memory traffic during long-context inference. Sparse attention is widely used to accelerate LLM inference by computing exact attention over a selected subset of tokens. To recover the contribution of tokens excluded from exact attention, recent methods apply coarse-grained compensation to the omitted attention tail. However, existing methods typically select tokens based on attention mass and only then compensate for the unselected tokens. This decoupled design overlooks their interaction: selection should prioritize tokens that would leave the largest compensation error if omitted. To address this limitation, we introduce CompKV, the first compensation-aware sparse attention framework that divides tokens into blocks and explicitly optimizes selection for the downstream compensation mechanism. Our theoretical analysis shows that the residual left by block-level mean compensation is governed by both block attention mass and within-block logit variation. We approximate this residual using compact block-level statistics, yielding a deployable selection criterion. We further develop an efficient asynchronous implementation. Experiments on RULER and LongBench-Pro show that CompKV performs best among the evaluated sparse baselines while delivering up to a $6.85\times$ self-attention speedup over full attention.

cs.LG↗

arXiv2026

Attend, Transform, or Silence: Operator-Level Visual Skipping for Efficient Multimodal LLM Inference

Multimodal large language models (MLLMs) increasingly process long visual-token sequences, increasing the overall inference computation. Existing acceleration methods usually remove visual tokens or skip visual-token updates in entire layers, but these coarse strategies may discard fine-grained evidence or suppress useful operators together with redundant ones. In this paper, we study visual-token computation from an answer-observable perspective and find that late visual-token updates can remain large while having little effect on answer-token representations. Motivated by this answer-silent redundancy, we decompose each Transformer layer into attention and FFN operators and show that useful visual computation is often operator-dominant and layer-dependent. We propose an operator-level visual-token skipping framework that preserves the full visual-token sequence while selectively bypassing redundant attention, FFN, or both. Experiments across three MLLM architectures and 10 VQA benchmarks show that our method achieves strong efficiency-accuracy trade-offs, reducing \textbf{33.7\%} TFLOPs on Qwen3-VL while retaining \textbf{99.5\%} of the vanilla model performance.

cs.CV↗

arXiv2026

RS-Prune: Training-Free Data Pruning at High Ratios for Efficient Remote Sensing Diffusion Foundation Models

Diffusion-based remote sensing (RS) generative foundation models are cruial for downstream tasks. However, these models rely on large amounts of globally representative data, which often contain redundancy, noise, and class imbalance, reducing training efficiency and preventing convergence. Existing RS diffusion foundation models typically aggregate multiple classification datasets or apply simplistic deduplication, overlooking the distributional requirements of generation modeling and the heterogeneity of RS imagery. To address these limitations, we propose a training-free, two-stage data pruning approach that quickly select a high-quality subset under high pruning ratios, enabling a preliminary foundation model to converge rapidly and serve as a versatile backbone for generation, downstream fine-tuning, and other applications. Our method jointly considers local information content with global scene-level diversity and representativeness. First, an entropy-based criterion efficiently removes low-information samples. Next, leveraging RS scene classification datasets as reference benchmarks, we perform scene-aware clustering with stratified sampling to improve clustering effectiveness while reducing computational costs on large-scale unlabeled data. Finally, by balancing cluster-level uniformity and sample representativeness, the method enables fine-grained selection under high pruning ratios while preserving overall diversity and representativeness. Experiments show that, even after pruning 85\% of the training data, our method significantly improves convergence and generation quality. Furthermore, diffusion foundation models trained with our method consistently achieve state-of-the-art performance across downstream tasks, including super-resolution and semantic image synthesis. This data pruning paradigm offers practical guidance for developing RS generative foundation models.

cs.CV↗