{"slug": "seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware", "title": "Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation", "summary": "Researchers discovered that Diffusion Multimodal Large Language Models reveal their valid semantic boundary at the first denoising step through a shift in MLP activation sparsity, enabling a training-free framework called Seer that truncates redundant computations. Seer accelerates inference by up to 31x while maintaining or improving performance on benchmarks like DocVQA.", "body_md": "arXiv:2607.14557v1 Announce Type: new\nAbstract: Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. Since the actual output length is unknown, output sequences are padded to a predefined maximum length, resulting in substantial redundant computation over unnecessary [EOS] tokens. In this work, we discover that DMLLMs implicitly reveal their valid semantic boundary at the very first denoising step through a distinct shift in MLP activation sparsity. Leveraging this observation, we propose Seer, a training-free framework that detects this boundary using a Signal-to-Noise Ratio (SNR)-based criterion and performs one-shot truncation of the redundant suffix for all subsequent computations. To preserve these theoretical gains during batched serving, Seer incorporates a hybrid execution strategy that maximizes throughput while seamlessly accommodating dynamic sequence lengths. Experimental results demonstrate that Seer effectively eliminates padding waste, accelerating throughput by up to $\\sim$31$\\times$. Across 9 benchmarks, Seer robustly maintains overall performance and even improves accuracy on complex visual tasks by mitigating noise leakage (e.g., DocVQA score increases from 63.52 to 63.66), offering a highly efficient, plug-and-play solution for DMLLM acceleration.", "url": "https://wpnews.pro/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware", "canonical_source": "https://www.machinebrief.com/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-wdsl", "published_at": "2026-07-17 04:00:00+00:00", "updated_at": "2026-07-17 05:02:59.926302+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-infrastructure"], "entities": ["Seer", "DocVQA"], "alternates": {"html": "https://wpnews.pro/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware", "markdown": "https://wpnews.pro/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware.md", "text": "https://wpnews.pro/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware.txt", "jsonld": "https://wpnews.pro/news/seeing-the-end-at-step-zero-accelerating-diffusion-mllms-via-mlp-sparsity-aware.jsonld"}}