Trang chủ / Tin tức / Khoa học & công nghệ
Khoa học & công nghệ

Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation

05/11/2026 · Thành viên: Ông. Nguyễn Đức Hải · Nguồn: OpenAlex

Diffusion Multimodal Large Language Models (DMLLMs) are highly effective for multimodal reasoning, yet their inference efficiency is significantly hindered by fixed-length generation constraints. S...

<p><strong>Tác giả:</strong> Qicheng Zhao, Qi Sun, Zheyu Yan</p>

<p>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 t...</p>

<p>📄 DOI: <a href="https://doi.org/https://doi.org/10.1145/3767308.3836260" target="_blank" rel="noopener">https://doi.org/10.1145/3767308.3836260</a></p>

<p>📖 <a href="https://doi.org/https://doi.org/10.1145/3767308.3836260" target="_blank" rel="noopener">Xem bài báo đầy đủ</a></p>

Bài viết được biên soạn lại bởi LIFVietnam. Nguồn tham khảo: OpenAlex