当前,采用下一 token 预测范式的自回归大型语言模型已经风靡全球,同时互联网上的大量合成图像和视频也早已让我们见识到了扩散模型的强大之处。
近日,MIT CSAIL 的一个研究团队(一作为 MIT 在读博士陈博远)成功地将全序列扩散模型与下一 token 模型的强大能力统合到了一起,提出了一种训练和采样范式:Diffusion Forcing(DF)。
![图片](https://image.jiqizhixin.com/uploads/editor/ab9fc883-1c21-4b3f-9054-6aec382306ea/640.png)
论文标题:Diffusion Forcing:Next-token Prediction Meets Full-Sequence Diffusion
论文地址:https://arxiv.org/pdf/2407.01392
项目网站:https://boyuan.space/diffusion-forcing
代码地址:https://github.com/buoyancy99/diffusion-forcing
如下所示,扩散强制在一致性和稳定性方面都明显胜过全序列扩散和教师强制这两种方法。
![图片](https://image.jiqizhixin.com/uploads/editor/f0f57219-5171-4b87-8862-d3f68938349d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2e04b4b5-049c-4e59-9126-1e008bdf07fc/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/f4cc6157-def9-4985-81ba-19ae27070cf9/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/a55fba97-7bb5-4157-8a50-75beefde1f08/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/af1353e1-3951-451a-912d-3cc07f2566c8/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/dd748e49-1ba7-4fe6-98e8-34295f6d463d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b789c729-3026-4abf-94c0-cebd242be5cc/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/b0e2a5be-bc7a-49c3-8856-7701079f526d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/8ea22383-2b17-4bdd-8c8f-0148c6108b55/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/3ef167a6-82f9-4572-b366-ef70e6474564/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/8bc24777-e0d1-4e3f-8e4a-51bf0d64a01d/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/cea46510-0f2c-400e-a221-30f83f54fe8b/640.png)
让自回归生成变得稳定 保持未来的不确定 长期引导能力
![图片](https://image.jiqizhixin.com/uploads/editor/b44b2cd7-59d3-4862-81dd-5e0162b5dd65/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/fd2beed1-ca9b-46a3-b66e-ef63808716ed/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/9bf4f46b-3cc3-4f96-906f-e718039f69d2/640.png)
具有灵活的规划范围 可实现灵活的奖励引导 能实现蒙特卡洛树引导(MCTG),从而实现未来不确定性
![图片](https://image.jiqizhixin.com/uploads/editor/e371cc3f-6606-46be-bde4-86351c6e60ed/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/12b4dc28-17fe-455d-959b-4fb812387474/640.png)
![图片](https://image.jiqizhixin.com/uploads/editor/2cb71f8b-2ec5-41c8-ac42-c22c12e5a285/640.gif)
![图片](https://image.jiqizhixin.com/uploads/editor/7e814775-8190-4218-a2f5-9325b7b30b07/640.png)
![](https://image.jiqizhixin.com/uploads/editor/15db91f8-0ece-4341-89a2-192a47c2de33/1721703517495.png)