The paper introduces Modality-Aware Expert Pruning (MAEP), a training‑free method that decomposes expert importance by modality in a single forward pass for Mixture‑of‑Experts multimodal large language models. MAEP uses a Delta‑H Output metric to measure the impact of removing an expert and leverages low‑attention vision tokens as a stable pruning signal, while preserving text‑side importance. Experiments on three MoE‑based MLLMs show that MAEP reduces the BF16 expert weight footprint by about 27 GB (≈47% of total memory), enabling a 58 GB model to fit on a single 40 GB GPU and outperforming modality‑agnostic baselines on most multimodal benchmarks.
arXiv:2608. 06901v1 Announce Type: cross Abstract: Vision-language models (VLMs) have achieved remarkable generalization across diverse multimodal tasks through large-scale pre-training, yet their rapidly increasing computational and memory requirements pose significant challenges for deployment in constrained environments.
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:2609.38823v1 Announce Type: new Abstract: Multimodal mixture-of-experts (MoE) models combine sparse expert activation with visual-language capabilities, yet their inference remains costly becau...
arXiv:2605. 05225v3 Announce Type: replace-cross Abstract: Mixture-of-Experts Multimodal Large Language Models (MoE MLLMs) suffer from a significant efficiency bottleneck during Expert Parallelism (EP) inference due to the straggler effect.
TokenSwap introduces a method to benchmark and reduce the modality gap in multimodal large language models (MLLMs) by interleaving semantically aligned images with text tokens. By transforming text-only benchmarks like MMLU into image‑interleaved versions (TokenSwap‑Bench), the authors reveal a pervasive performance drop—averaging 19.6%—across 42 MLLMs, with reasoning models showing smaller gaps. Training with TokenSwap mitigates this gap while maintaining strong text‑only and vision‑language performance.
VisionWeave introduces elastic visual representation weaving, a native capability for multimodal large language models that learns where and at what granularity to encode visual information. The method combines a gated spatial pooler for coarse representations with a granularity router that allocates content‑adaptive token usage, trained end‑to‑end on large‑scale data. Experiments on Qwen3.5‑4B and Qwen3.8‑27B show that VisionWeave can save 43.0% of tokens while preserving 98.9% of performance across eight benchmarks, and delivers significant throughput gains and latency reductions when deployed on the SGLang serving engine.
SinkPruner is a training‑free framework that prunes visual tokens for multimodal large language models by first removing high‑norm redundant tokens with a visual sanitizer and then selectively keeping tokens that align with the text query using a text‑guided pruner. The coarse‑to‑fine design reduces attention sink and dispersion, enabling an 89% token reduction while preserving 96.5% of LLaVA‑1.5’s performance and 91.8% of Qwen2.5‑VL’s performance across twelve image‑language and four video‑language benchmarks. The visual sanitizer also improves existing pruning methods, showing strong transferability.
arXiv:2603. 18492v3 Announce Type: replace Abstract: Mixture-of-Experts (MoE) language models increase parameter capacity without proportional per-token computation, yet deployment still requires storing the full expert pool, making expert pruning important for reducing memory and serving overhead.
arXiv:2609.37515v1 Announce Type: cross Abstract: Thorough evaluation of vision-language models (VLMs) has become prohibitively expensive, as benchmarks span an ever-broader spectrum of capabilities...
arXiv:2608.06411v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) achieve strong performance across diverse vision-language tasks, but their efficiency is limited by...
The paper investigates whether the sparsity of Mixture-of-Experts (MoE) models leads to intrinsic semantic organization across modalities and domains. It shows that experts naturally specialize semantically even without explicit modular training. The authors propose ExpertLens, a data‑free method that decodes router weights to identify domain‑specialized experts, enabling selective fine‑tuning that matches or exceeds full fine‑tuning while updating only 21.7–47.0% of parameters and achieving a 4.0× speedup, outperforming LoRA in both performance and efficiency.
The paper investigates how to scale a Transformer-based vision‑language model trained from scratch on multimodal inputs under a fixed compute budget. It finds that minimal loss follows a predictable compute law while optimal model size and token count obey power‑law scaling, with language and multimodal objectives showing distinct allocation trends. The study also proposes a budget‑compensation rule that allows additional multimodal tokens to offset text‑efficiency penalties, and demonstrates that native multimodal pre‑training improves spatial reasoning and few‑shot learning.