Загрузка...

ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding (Dec 2025)

Title: ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding (Dec 2025)
Link: http://arxiv.org/abs/2512.13586v1
Date: December 2025

Summary:
This paper introduces ReFusion, a novel masked diffusion model designed to overcome the efficiency and coherence limitations of existing parallel decoding methods. ReFusion employs a slot-based "plan-and-infill" strategy where a diffusion process first identifies weakly dependent slots, followed by parallel autoregressive decoding of those slots. A key architectural innovation allows ReFusion to achieve full Key-Value (KV) cache reuse—typically impossible in bidirectional diffusion models—by dynamically reordering generated slots. Extensive experiments show ReFusion outperforms prior masked diffusion models by 34% while being 18x faster, and surpasses strong autoregressive models like Qwen3-8B in performance with a 2.33x speedup.

Key Topics:
- Masked Diffusion Models
- Parallel Decoding
- Autoregressive Models
- KV Cache Reuse
- Plan-and-Infill
- Slot-based Generation
- Large Language Models

Chapters:
00:00 - Introduction to ReFusion
01:29 - Inference Latency Bottlenecks
03:07 - MDM Coherence Challenges
04:15 - The Slot-Based Architecture
05:35 - Achieving Full KV Reuse
07:26 - Step 1: Slot Planning
08:08 - Step 2: Verification and Infilling
09:58 - Benchmarking and Robustness
11:40 - Non-Linear Coding Example
12:45 - Limitations and Future Outlook

Stock video credits:
- tunnel motions - https://www.pexels.com/@tunnelmotions
- Pixabay - https://www.pexels.com/@pixabay
- Bedrijfsfilmspecialist.nl - https://www.pexels.com/@bedrijfsfilmspecialist-nl-1284006
- Colin Jones - https://www.pexels.com/@larchmedia
- Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
- Dan Cristian Pădureț - https://www.pexels.com/@paduret
- Ron Lach - https://www.pexels.com/@ron-lach
- cottonbro studio - https://www.pexels.com/@cottonbro
- Silviu Din - https://www.pexels.com/@silviu-din-1620549
- Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
- Pressmaster - https://www.pexels.com/@pressmaster
- Kindel Media - https://www.pexels.com/@kindelmedia
- Kelly - https://www.pexels.com/@kelly
- crazy motions - https://www.pexels.com/@crazy-motions-80195021
- Charlie Mounsey - https://www.pexels.com/@charlie-mounsey-1653902
- Colors Motion Graphics - https://www.pexels.com/@colors-motion-graphics-183847699
- StefWithAnF - https://www.pexels.com/@stefwithanf-1955763
- Trippy Lagoon - https://www.pexels.com/@trippy-lagoon-511515544
- Soumya - https://www.pexels.com/@soumya-1446957
- José Alfredo Munguía Lira - https://www.pexels.com/@rectorretro
- Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
- Anete Lusina - https://www.pexels.com/@anete-lusina
- Stas Knop - https://www.pexels.com/@stasknop
- Pachon in Motion - https://www.pexels.com/@pachon-in-motion-426015731

Видео ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding (Dec 2025) канала AI Paper Slop
Яндекс.Метрика
Все заметки Новая заметка Страницу в заметки
Страницу в закладки Мои закладки
На информационно-развлекательном портале SALDA.WS применяются cookie-файлы. Нажимая кнопку Принять, вы подтверждаете свое согласие на их использование.
О CookiesНапомнить позжеПринять