1
MiniMax-Music3 音乐模型发布：AI 生成最长 5 分钟歌曲
MiniMax-Music3 强势登场，分层自回归架构生成最长5分钟完整歌曲，技术细节全解析。
IT之家 8 月 14 日消息，AI 公司 MiniMax 今天（8 月 14 日）推出音乐生成模型 MiniMax-Music3， 可根据歌词和音乐描述生成最长 5 分钟的完整歌曲。 架构方面，MiniMax-Music3 模型采用分层自回归架构： Global LLM（全局语言模型）参数规模为 …
2
Answer First, Reason Later: Commitment Order in Diffusion LLMs
扩散LLM自由排序承诺反而成推理软肋，揭秘答案先行的失败机制
arXiv:2608.05687v1 Announce Type: cross Abstract: Masked diffusion language models (dLLMs) can commit tokens in any order -- a freedom marketed as the…
3
Autoregressive Modeling of Film with Applications in Video Montage
自回归模型也能拍电影？这篇论文教你用AI自动生成视频蒙太奇，颠覆传统剪辑思路。
arXiv:2607.14645v1 Announce Type: new Abstract: This work introduces FilmGPT, an autoregressive transformer designed to address the challenge of video…
4
Generative Lane Topology Reasoning via Autoregressive Model with Geometry Prior
用自回归模型融合几何先验，车道拓扑推理更精准高效。
arXiv:2606.31814v1 Announce Type: new Abstract: Lane topology reasoning aims to construct a lane graph from onboard sensor observations. Existing meth…
5
GEAR: Guided End-to-End AutoRegression for Image Synthesis
GEAR 打破两阶段训练，用端到端自回归协同优化图像生成，让 tokenizer 与生成器无缝配合，产出的图像更真实。
arXiv:2606.32039v1 Announce Type: new Abstract: Visual generative models are typically trained in two stages. A tokenizer is first trained for reconst…
6
独家｜获超亿美元融资，Sand.ai 曹越：为什么视频是通往世界模型最重要的路径
Sand.ai获超亿美元融资，曹越详解为何视频是通往世界模型的关键路径
“每一代模型，我们都在押注一个非共识。” 文｜邓咏仪 编辑｜张雨忻 Sand.ai 创始人曹越，不太关心自己站在共识的哪一边。 Sand.ai 是一家视频生成模型和产品公司，成立于2024年1月。曹越创立Sand.ai 的故事也已经被讲过很多遍：在上一段创业“光年之外”戛然而止后，曹越很快就投入到 …
7
HiFiVe: High-Fidelity Vehicle Generation Leveraging Auto-Regressive 2D Generative Priors
车载3D生成新方案，借助2D自回归先验实现高保真车辆重建
arXiv:2606.25300v1 Announce Type: new Abstract: Existing 3D vehicle generation methods often suffer from low geometric fidelity and blurry textures, h…
8
Logit Distance Bounds Representational Similarity
理论揭示：判别模型的条件分布等价性如何约束内部表示的唯一性，为理解模型表征相似性提供新视角
arXiv:2602.15438v3 Announce Type: replace-cross Abstract: For a broad family of discriminative models that includes autoregressive language models, id…
9
DREAM: Dense Retrieval Embeddings via Autoregressive Modeling
自回归建模加持，让密集检索嵌入更智能，刷新稠密向量生成思路
arXiv:2606.24667v1 Announce Type: new Abstract: Dense retrieval embedding models are a fundamental component of modern retrieval-based AI systems. Mos…
10
EfficientRollout: System-Aware Self-Speculative Decoding for RL Rollouts
针对强化学习训练中 rollout 生成缓慢的瓶颈，提出了一种系统感知的自推测解码方法，显著加速 LLM 推理。
arXiv:2606.18967v1 Announce Type: new Abstract: Reinforcement learning (RL) has become a representative post-training paradigm for LLMs, enabling stro…
11
Learning from the Self-future: On-policy Self-distillation for dLLMs
探索将自蒸馏技术从自回归模型迁移至扩散LLM，突破传统方法限制，为后训练提供新思路。
arXiv:2606.18195v1 Announce Type: new Abstract: On-policy self-distillation (OPSD) has proven effective for post-training large language models (LLMs)…
12
Sumi: Open Uniform Diffusion Language Model from Scratch
基于均匀扩散的语言模型，支持任意位置同时更新token，实现比自回归模型更灵活的文本生成
arXiv:2606.19005v1 Announce Type: cross Abstract: Diffusion models have become a promising alternative to autoregressive models. Among these, uniform …
13
Shattering the Autoregressive Curse: Dynamic Epistemic Entropy Orchestrated Erasable Reinforcement Learning for LLMs
新方法通过动态认知熵和可擦除RL打破LLM自回归诅咒，实现长程逻辑推理优化。
arXiv:2606.17735v1 Announce Type: new Abstract: Although reinforcement learning (RL) has expanded the cognitive boundaries of large language models (L…
14
DICE: Diffusion Large Language Models Excel at Generating CUDA Kernels
扩散大语言模型在CUDA内核生成任务上超越自回归模型，新论文DICE展示显著优势。
arXiv:2602.11715v2 Announce Type: replace Abstract: Diffusion large language models (dLLMs) have emerged as a compelling alternative to autoregressive…
15
GPT-Based Fast Simulation of CLAS12 Detector Hits via Conditional Autoregressive Generation
用GPT做物理探测器模拟？论文提出条件自回归生成方法，加速CLAS12探测器响应仿真，AI与高能物理的跨界创新。
arXiv:2606.16035v1 Announce Type: cross Abstract: Modern particles physics experiments have demonstrated an increasing need for fast, high-fidelity de…
16
Residual Context Diffusion Language Models
扩散语言模型借助残差上下文实现并行多token解码，突破自回归限制，大幅提升文本生成效率
arXiv:2601.22954v2 Announce Type: replace-cross Abstract: Diffusion Large Language Models (dLLMs) have emerged as a promising alternative to purely au…
17
Google's latest DiffusionGemma open AI model comes with a 4x speed boost
Google开源DiffusionGemma，非自回归文本模型，速度提升4倍，类似图像去噪生成。
Diffusion AI is most common in image generation, but it can make text outputs much faster.
18
MAGE: All-[MASK] Block Already Knows Where to Look in Block Diffusion LLM
块扩散语言模型MAGE让所有[MASK]块自主定位关键位置，显著提升生成速度与连贯性。
arXiv:2602.14209v2 Announce Type: replace Abstract: Block diffusion LLMs are an emerging paradigm for parallel language generation, but their KV cachi…
19
When Autoregressive Consistency Hurts Safety Alignment
自回归一致性让大模型安全对齐脆弱不堪：微调只能重塑输出开头几个token，后续轨迹难以纠正。
arXiv:2606.04168v1 Announce Type: new Abstract: Safety alignment in large language models (LLMs) is fragile in part because it is often shallow: fine-…
20
Compositional Generalization in Autoregressive Models via Logit Composition
新方法通过logit组合提升自回归模型的组合泛化能力，突破结构泛化瓶颈。
arXiv:2605.28304v1 Announce Type: new Abstract: Composing autoregressive models remains a core challenge in understanding how large language models ca…