TwinPrune introduces a unified spatial‑density signal to score token importance for both inference phases of vision‑language models: a bipartite merge during prefill and KV eviction during decode. By separating benchmarks into prefill‑only and full‑pipeline probes, the authors show that decode‑phase pruning must be evaluated on full‑pipeline tasks, as prefill probes can be misleading. Experiments on seven model variants demonstrate that density‑based prefill compression preserves accuracy while scaling linearly with visual token count, and that density‑top‑k KV eviction outperforms strong baselines on TextVQA and MM‑Vet across multiple retention ratios, achieving a Pareto frontier in compute, memory, and accuracy.
arXiv:2610.33688v1 Announce Type: new Abstract: Visual token pruning substantially reduces the inference cost of multimodal large language models (MLLMs), yet existing methods are mostly training-fre...
arXiv:2605.10980v2 Announce Type: replace Abstract: Diffusion Language Models (dLLMs) have garnered significant attention for their potential in highly parallel processing. The parallel capabilities...
arXiv:2610.03389v1 Announce Type: cross Abstract: Vision language models (VLMs) incur substantial inference cost because every visual token is processed by the attention and MLP projections of every...
MWOP (Modality-aware Width-wise Operation Pruning) is a method that independently prunes visual‑to‑visual, text‑to‑visual, and text‑to‑text attention paths within each layer of multimodal large language models, and separately selects feed‑forward network channels for visual and textual inputs. It uses a first‑order Taylor criterion to guide pruning, re‑evaluates FFN importance after attention pruning, and applies LoRA‑based recovery training. The approach is paired with path‑sparse Triton attention kernels and compact visual‑side FFN execution to achieve practical acceleration, preserving token sequences while reducing computation. "whyItMatters":"MWOP achieves a 1.6× prefill speedup on LLaVA‑OneVision‑7B while retaining 99.7% performance, and further boosts token‑compression methods to 2.9× and 2.7× speedups, demonstrating its effectiveness across architectures."
arXiv:2609.05916v1 Announce Type: cross Abstract: Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose subs...
arXiv:2607. 28418v1 Announce Type: cross Abstract: Pruning is a promising approach for improving the efficiency of LLMs.
arXiv:2609.35232v2 Announce Type: replace-cross Abstract: Visual-token compression is effective for improving the efficiency of vision-language models, but under extreme compression budgets, token pr...
The paper introduces PACE, a Pareto‑adaptive compression framework designed for native multimodal large language models (MLLMs) that directly tokenize pixel patches without external vision encoders. By analyzing token discriminability and heterogeneous redundancy across spatial dimensions, PACE decomposes attention into height‑ and width‑specific saliency maps and applies a Pareto‑optimization strategy to allocate token budgets, achieving near‑lossless performance with only 25‑50% of visual tokens and more than 2× FLOP savings. Experiments demonstrate that PACE not only preserves but sometimes improves benchmark results compared to existing ViT‑based compression methods.
arXiv:2603. 00198v2 Announce Type: replace-cross Abstract: Token reduction accelerates long-video vision--language models (VLMs), but existing methods target Transformers, where reduction is treated as token pruning.
arXiv:2609.26333v1 Announce Type: new Abstract: Prefill and decode reward different approaches to quantization: low-precision arithmetic accelerates prompt processing, while compact weights reduce me...
The paper demonstrates that greedy decoding from large language models is not precision‑invariant: the same model, prompt, and decoding algorithm can produce different outputs when run in BF16 versus FP16 on identical hardware. Across six models (1.1B–7B parameters, four families, and 12B) and three benchmarks, 49–100 % of prompts diverge, with a single token flip often cascading into trajectory‑level divergence. The authors develop an empirical error‑propagation analysis that identifies the top‑two logit margin at the LM head as the key factor, and they propose a low‑overhead intervention—selective FP32 LM head recomputation—that improves exact agreement by 22–36 percentage points with less than 4 % latency overhead. "whyItMatters":"The findings reveal that precision choices can fundamentally alter model outputs, challenging the assumption of deterministic greedy decoding and highlighting the need for precision‑aware inference strategies."
CatchUp is a training‑free framework that improves efficiency in large vision‑language models by depth‑aware re‑admission of visual tokens. It repeatedly evaluates query‑conditioned relevance, processes only selected tokens per decoder interval, and preserves deferred tokens for later re‑admission, using reference‑to‑current transition signals to approximate missed representation evolution. Experiments on three referring‑expression grounding datasets and six multimodal benchmarks show that CatchUp retains over 92% of full‑model performance with 88.9% fewer active tokens and reduces end‑to‑end latency by 8.0% at a 64‑token budget.