作者 | 李崇轩 编辑 | 自动驾驶之心 原文链接： https://www.zhihu.com/question/1908479621466396378/answer/1910672718174589774 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近30个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 大家好，我是中国人民大学高瓴人工智能学院李崇轩，因为做的非常相关，来回答一下这个问题。 我在连续扩散模型和朱军老师以及师弟师妹们有很多合作，代表性工作有 Analytic-DPM，U-ViT, DPM-Solver，ProlificDreamer，DPM-Solver++，unidiffuser 等 等。 我在人大的课题组很年轻，组内在离散扩散模型的代表性工作有 RADD，Scaling Law for MDM，LLaDA，LLaDA-V 和这两天即将发布的 LLaDA 1.5。 我想可以按照时间划分为两个阶段来介绍一下这个领域，然后发表一下我的看法。 第一阶段：2022-2024年底，扩散语言模型偏基础研究的阶段 ...

Core Insights - The article discusses the rapid advancements in diffusion large language models (LLMs), highlighting their potential as strong competitors to traditional LLMs [2][7] - A recent paper from a collaborative research team proposes an efficient decoding strategy combined with reinforcement learning for masked diffusion large language models (MDLM), significantly improving their reasoning performance and efficiency [2][21] Group 1: Problem Identification - Masked diffusion large language models like LLaDA exhibit capabilities comparable to autoregressive models but face challenges with full diffusion-style decoding, which is less effective than block-wise decoding [7][9] - The decoding process of MDLMs often encounters an issue where early generation of <EOS> tokens leads to performance degradation, creating a decoding trap [14][15] Group 2: Proposed Solutions - The research team introduces an early rejection mechanism for <EOS> tokens to suppress their confidence during early decoding steps, thus preventing premature termination of generation [15] - A power-increasing decoding step scheduler is designed to optimize the decoding process, reducing the inference steps from O(L) to O(logL), thereby accelerating reasoning [15][16] Group 3: Consistency Trajectory Optimization - The team proposes a consistency trajectory grouping strategy (CJ-GRPO) to address inconsistencies between rollout and optimization trajectories, enhancing training stability and effectiveness [16] - By combining the early rejection mechanism, increasing step scheduler, and CJ-GRPO, the model can maintain performance comparable to baseline methods while significantly reducing decoding steps [16][24] Group 4: Experimental Results - Extensive experiments demonstrate that the proposed methods outperform baseline models in mathematical reasoning and planning tasks, with performance improvements of up to 2-4 times in certain benchmarks [23][24] - The results indicate that the combination of CJ-GRPO with EOSER and ASS maintains competitive performance in low-step inference scenarios, achieving a balance of speed and quality [24] Group 5: Future Directions - The article suggests exploring hybrid reasoning modes that combine the strengths of diffusion and autoregressive models to meet diverse task requirements [26]

Core Insights - The article discusses the limitations of the traditional left-to-right sampling method in large language models and introduces the Masked Diffusion Language Model (MDLM) as a potential alternative, highlighting its advantages in various tasks [1][5][15]. Group 1: MDLM Advantages - MDLM allows for arbitrary order decoding and supports multi-token parallel decoding, which can enhance performance in certain tasks like Sudoku [1][4]. - However, recent findings indicate that in mathematical and coding tasks, arbitrary order algorithms often perform worse than left-to-right sampling, and multi-token decoding significantly reduces performance [1][4][5]. - The study suggests that using MDLM for left-to-right sampling can be an efficient approach for reasoning and coding tasks, especially when block sizes are enforced to maintain a semi-autoregressive structure [3][5]. Group 2: Prompting and In-filling Techniques - The researchers propose a new prompting method called "prompting-as-in-filling," which allows users to add context at multiple positions rather than just the beginning of the sequence [6][18]. - They introduce a "reasoning-as-in-filling" framework that utilizes a reasoning template to guide the model in generating reasoning trajectories based on a given budget and format [6][18][19]. Group 3: Early Exit Mechanism - The "reasoning-as-in-filling" method enables the model to quantify answer uncertainty during the reasoning process, allowing for early exits to reduce computational costs [8][19]. - For instance, in the GSM8k dataset, this approach led to a 24% reduction in function calls without sacrificing accuracy [8]. Group 4: Multi-Token Entropy Decoding (MED) - The researchers developed an adaptive multi-token decoder called MED, which only performs parallel decoding when the conditional entropy of the additional positions is below a set threshold, thus controlling the deviation from single-token decoding [10][24]. - Experimental results show that the MED method can achieve a 2-3 times reduction in function calls while maintaining performance [11][26]. Group 5: Post-Training Capabilities - The study highlights that MDLM's in-filling capabilities unlock new sampling and post-training mechanisms, allowing for effective post-training without the need for complex prompt designs or additional models [22][23]. - By sampling reasoning trajectories from the posterior distribution, researchers can enhance the model's performance on reasoning tasks [22][23][33]. Group 6: Performance Metrics - The article presents various performance metrics, showing that using MED can lead to significant speed improvements while maintaining accuracy across different datasets [26][30]. - The results indicate that early exit mechanisms combined with MED can further optimize computational efficiency, particularly in the LLaDA model [31][32].

Core Insights - Ant Group has officially announced the open-source release of dInfer, the industry's first high-performance inference framework for diffusion language models [1][5] - dInfer demonstrates a significant improvement in inference speed, achieving a 10.7 times increase compared to NVIDIA's Fast-dLLM framework, and reaching a speed of 1011 tokens per second in the HumanEval code generation task [1][4] - The framework addresses key challenges in diffusion language model inference, including high computational costs, KV cache failures, and parallel decoding [1][2] Summary by Sections - **Performance Metrics** - dInfer achieves an average inference speed of 681 tokens per second, compared to 63.6 tokens per second for Fast-dLLM, marking a 10.7 times improvement [4] - When compared to the AR model Qwen2.5-3B, dInfer's average inference speed is 2.5 times faster, at 681 tokens per second versus 277 tokens per second [5] - **Technical Architecture** - dInfer is designed with a modular architecture that includes four core components: Model, KV-Cache Manager, Iteration Manager, and Decoder, allowing developers to customize and optimize their configurations [2] - Each module integrates targeted solutions to overcome the three main challenges faced by diffusion language models [2] - **Industry Impact** - The launch of dInfer signifies a critical step in transitioning diffusion language models from theoretical feasibility to practical efficiency, connecting cutting-edge research with industrial applications [5] - Ant Group invites global developers and researchers to explore the potential of diffusion language models, aiming to build a more efficient and open AI ecosystem [5]

Core Viewpoint - The article discusses the current state of large language models (LLMs) and the challenges they face in achieving Artificial General Intelligence (AGI), emphasizing the need for new paradigms beyond the existing autoregressive (AR) models [4][10][18]. Group 1: Current Challenges in AI Models - Ilya, a prominent AI researcher, warns that data extraction has reached its limits, hindering the progress towards AGI [2][4]. - The existing LLMs often exhibit significant performance discrepancies, with some capable of outperforming human experts while others struggle with basic tasks [13][15]. - The autoregressive model's limitations include a lack of bidirectional modeling and the inability to correct errors during generation, leading to fundamental misunderstandings in tasks like translation and medical diagnosis [26][27][18]. Group 2: New Directions in AI Research - Elon Musk proposes a "purified data" approach to rewrite human knowledge as a potential pathway to AGI [5]. - Researchers are exploring multimodal approaches, with experts like Fei-Fei Li emphasizing the importance of visual understanding as a cornerstone of intelligence [8]. - A new paradigm, the diffusion model, is being introduced by young scholars, which contrasts with the traditional autoregressive approach by allowing for parallel decoding and iterative correction [12][28]. Group 3: Development of LLaDA-MoE - The LLaDA-MoE model, based on diffusion theory, was announced as a significant advancement in the field, showcasing a new approach to language modeling [12][66]. - LLaDA-MoE has a total parameter count of 7 billion, with 1.4 billion activated parameters, and has been trained on approximately 20 terabytes of data, demonstrating its scalability and stability [66][67]. - The model's performance in benchmark tests indicates that it can compete with existing autoregressive models, suggesting a viable alternative path for future AI development [67][71]. Group 4: Future Prospects and Community Involvement - The development of LLaDA-MoE represents a milestone in the exploration of diffusion models, with plans for further scaling and improvement [72][74]. - The team emphasizes the importance of community collaboration in advancing the diffusion model research, similar to the development of autoregressive models [74][79]. - Ant Group's commitment to investing in AGI research reflects a strategic shift towards exploring innovative and potentially high-risk areas in AI [79].

Core Viewpoint - The article discusses the development of the LLaDA-MoE model, the first native MoE architecture diffusion language model trained from scratch, which demonstrates significant performance and efficiency advantages over traditional autoregressive models [2][15][18]. Group 1: Model Development and Performance - The LLaDA-MoE model was trained on 20 terabytes of data and features 1.4 billion active parameters, achieving performance comparable to denser autoregressive models like Qwen2.5-3B while maintaining faster inference speeds [15][17][29]. - The LLaDA series has rapidly evolved, with LLaDA-MoE being a notable milestone, surpassing previous models like LLaDA1.0/1.5 and Dream-7B in various benchmark tests [13][18][29]. - The model's architecture allows for significant scaling potential, with plans to explore higher sparsity ratios and larger MoE diffusion language models [29][40]. Group 2: Technical Innovations and Advantages - The diffusion model approach allows for parallel decoding, bidirectional modeling, and iterative correction, addressing limitations of autoregressive models such as serial bottlenecks and lack of error correction capabilities [38][40]. - Evidence suggests that diffusion language models can achieve better learning outcomes than autoregressive models, particularly in scenarios with limited data, demonstrating a data utilization efficiency that can exceed three times that of autoregressive models [40][41]. - The training framework and infrastructure developed by Ant Group, including the ATorch framework, supports the efficient training of large-scale MoE models [25][26]. Group 3: Strategic Vision and Future Directions - The development of LLaDA-MoE reflects a strategic choice to explore high-potential areas in AI, moving beyond established paths to enhance the limits of intelligence [44][47]. - Ant Group's commitment to innovation is evident in its previous projects and ongoing research in areas like dynamic MoE architectures and hybrid linear architectures, all aimed at achieving general artificial intelligence (AGI) [45][46][47].

Core Viewpoint - The article discusses the advancements in Vision-Language Models (VLMs) and introduces LLaDA-VLA, the first Vision-Language-Action Model developed using large language diffusion models, which demonstrates superior multi-task performance in robotic action generation [1][5][19]. Group 1: Introduction to LLaDA-VLA - LLaDA-VLA integrates Masked Diffusion Models (MDMs) into robotic action generation, leveraging pre-trained multimodal large language diffusion models for fine-tuning and enabling parallel action trajectory prediction [5][19]. - The model architecture consists of three core modules: a vision encoder for RGB feature extraction, a language diffusion backbone for integrating visual and language information, and a projector for mapping visual features to language token space [10][7]. Group 2: Key Technical Innovations - Two major breakthroughs are highlighted: - Localized Special-token Classification (LSC), which reduces cross-domain transfer difficulty by classifying only action-related special tokens, thus improving training efficiency [8][12]. - Hierarchical Action-Structured Decoding (HAD), which explicitly models hierarchical dependencies between actions, resulting in smoother and more reasonable generated trajectories [9][13]. Group 3: Performance Evaluation - LLaDA-VLA outperforms state-of-the-art methods across various environments, including SimplerEnv, CALVIN, and real robot WidowX, achieving significant improvements in success rates and task completion metrics [4][21]. - In specific task evaluations, LLaDA-VLA achieved an average success rate of 58% across multiple tasks, surpassing previous models [15]. Group 4: Experimental Results - The model demonstrated a notable increase in task completion rates and average task lengths compared to baseline models, validating the effectiveness of the proposed LSC and HAD strategies [18][14]. - In a comparative analysis, LLaDA-VLA achieved a success rate of 95.6% in a specific task, significantly higher than other models [14][18]. Group 5: Research Significance and Future Directions - The introduction of LLaDA-VLA establishes a solid foundation for applying large language diffusion models in robotic operations, paving the way for future research in this domain [19][21]. - The design strategies employed in LLaDA-VLA not only enhance model performance but also open new avenues for exploration in the field of embodied intelligence [19].

Group 1 - The article discusses the potential of Diffusion LLMs, particularly Gemini Diffusion, as a significant breakthrough in AI, challenging traditional autoregressive models [3][4][5] - Gemini Diffusion demonstrates high generation efficiency, achieving an average sampling speed of 1479 TPS and up to 2000 TPS in encoding tasks, outperforming Gemini 2.0 Flash-Lite by 4-5 times [4][6] - The parallel generation mechanism of the diffusion architecture allows for efficient processing, which could lead to reduced computational costs compared to autoregressive models [6][7] Group 2 - Mary Meeker emphasizes that the speed of AI development surpasses that of the internet era, highlighting the cost disparity between AI model training and inference [1][2] - The article suggests that the rise of open-source models in China may impact the global supply chain, indicating a shift in competitive dynamics within the industry [1][2] - The balance between computational investment and commercial returns is crucial for enterprises as AI inference costs decline [1][2]

Core Viewpoint - The article discusses the advancements in diffusion language models (dLLMs), particularly focusing on Google's Gemini Diffusion and its implications for AI development, highlighting the speed and performance improvements over traditional autoregressive models [1][8][35]. Group 1: Gemini Diffusion and Its Features - Gemini Diffusion is noted for its impressive generation speed, being five times faster than previous models, and its ability to handle programming tasks effectively [2][8]. - The underlying mechanism of diffusion models allows for rapid iteration and error correction during the generation process, distinguishing it from autoregressive models [2][3]. - Gemini Diffusion's sampling speed can reach an astonishing 1479 tokens per second, showcasing its potential in various benchmarks [8][9]. Group 2: Development of Diffusion Language Models - Prior to Gemini Diffusion, several research teams explored the feasibility of diffusion-based LLMs, including Stanford's Diffusion-LM and Fudan University's DiffusionBERT [3][4]. - The introduction of LLaDA, the first 8 billion parameter diffusion language model, marked a significant milestone in the field, achieving performance comparable to LLaMA 3 [4][21]. - Following LLaDA, other models like d1 and LaViDa have emerged, further establishing LLaDA as a foundational model in dLLM research [20][21]. Group 3: Multimodal Diffusion Language Models - The emergence of diffusion multimodal language models (dMLLMs) is highlighted, with LLaDA-V and MMaDA being prominent examples that integrate visual and language processing capabilities [10][31]. - LLaDA-V combines visual instruction fine-tuning with the diffusion mechanism, demonstrating strong performance in multimodal understanding tasks [26][27]. - MMaDA showcases innovations in text reasoning and multimodal understanding, solidifying its position as a leading research outcome in the dMLLM space [31][32]. Group 4: Future Directions and Implications - The article emphasizes the shift from autoregressive models to diffusion models as a significant paradigm change in AI, suggesting broader implications for future research and applications [35][36]. - The ongoing evolution of models like LLaDA and Gemini Diffusion indicates a growing ecosystem around dLLMs and dMLLMs, with potential applications extending into quantum computing [35][36].