综合 Survey 论文索引

最后更新: 2026-04-16 | 40+ 篇论文, 9 大分类

全面综述

• Understanding World or Predicting Future? (ACM CSUR 2025, arXiv:2411.14499) — 理解vs预测二元分类

• Simulating the Visual World with AI: A Roadmap (arXiv:2511.08585, 2025) — 四代视频模型演进路线图

• 3D and 4D World Modeling: A Survey (arXiv:2509.07996, 2025) — VideoGen/OccGen/LiDARGen 三轴分类

• Is Sora a World Simulator? (arXiv:2405.03520, 2024) — 从 Sora 出发的通用 WM 综述

具身 AI 专题

• A Comprehensive Survey on WM for Embodied AI (arXiv:2510.16732, 2025) — 功能×时间×空间三轴分类

• A Survey of Embodied World Models (清华/Frontiers, 2026) — 架构/训练/应用/评估四维度

• Learning Embodied Intelligence from Simulators and WMs (arXiv:2507.00917, 2025) — sim-to-real 挑战

• Large Model Empowered Embodied AI (arXiv:2508.10399, 2025) — LLM/VLM 赋能具身决策

• Video Generation Models in Robotics (arXiv:2601.07823, 2026) — 视频模型在机器人的应用

• Physically Embodied 3DGS for Robotics (arXiv:2406.10788, 2024) — 3DGS + 物理属性实时世界模型

自动驾驶专题

• World Models for AD: An Initial Survey (arXiv:2403.02622, 2024) — 首批 AD WM 系统综述

• A Survey of WM for AD (arXiv:2501.11260, 2025) — 三层分类: 生成/规划/闭环

• Exploring Video Gen and WMs in AD (arXiv:2411.02914, 2024) — 扩散模型在场景模拟

• The Role of WMs in Shaping AD (arXiv:2502.10498, 2025) — 按预测模态分类

3D/4D 专题

• 3D and 4D World Modeling (arXiv:2509.07996, 2025) — 首篇 3D/4D 专题综述

• A Survey on 3D Gaussian Splatting (arXiv:2401.03890, 2024) — 3DGS 全景

视频生成→世界模型

• How Far is Video Gen from World Model (arXiv:2411.02385, 2024) — 物理定律视角评测

• Survey of Video Diffusion Models (arXiv:2504.16081, 2025) — 扩散视频生成全景

• Pandora: Towards General WM (arXiv:2406.09455, 2024) — 自回归+扩散混合架构

天气/气候基础模型

• Deep Learning and Foundation Models for Weather (arXiv:2501.06907, 2025) — 三类训练范式

• ClimaX (ICML 2023) — 首个通用气候基础模型

• GraphCast (Science 2023) — 图神经网络全球天气预报

• Pangu-Weather (Nature 2023) — 3D 变换器超越数值预报

评估基准

• WorldModelBench (arXiv:2502.20694, CVPR 2025 Workshop) — 6.7万人工标注, SOTA 仅 61%

• WorldScore (arXiv:2504.00983, ICCV 2025) — 统一世界生成基准

• WorldBench (arXiv:2601.21282, 2026) — 概念级物理评估

• PhysBench (ICLR 2025) — 75 个 VLM, 10002 条目

• Revisiting Othello WM Hypothesis (ICLR 2025) — 七模型 99% 无监督 grounding

批评与反思

• How Far is Video Gen from WM (arXiv:2411.02385, 2024) — 视频模型违反物理定律

• Stochastic Parrot on LLM's Shoulder (arXiv:2502.08946, ICLR 2025) — LLM 物理概念落后人类 40%

• On the Dangers of Stochastic Parrots (FAccT 2021, Bender et al.) — "随机鹦鹉"批评

• OthelloGPT: Bag of Heuristics (LessWrong, 2024) — 局部规则集合非统一世界模型

• LLMs and World Models Part 1+2 (Mitchell, 2025) — 托勒密本轮 vs 太阳系仪

技术基础 (奠基论文)

• World Models (NeurIPS 2018, Ha & Schmidhuber) — "梦境"中训练策略

• Dreamer V1-V3 (ICLR 2020/2021, Nature 2023) — RSSM + 想象训练

• A Path Towards Autonomous Machine Intelligence (LeCun 2022) — JEPA 框架蓝图

• VAE (ICLR 2014, Kingma & Welling) — ELBO + 重参数化

• DDPM (NeurIPS 2020, Ho et al.) — 去噪扩散开创

• OthelloGPT (ICLR 2023, Li et al.) — 涌现世界表示

• Linear Representations (BlackboxNLP 2023, Nanda) — 线性探针 + 因果干预

• Space and Time Neurons (ICLR 2024, Gurnee & Tegmark) — 空间/时间线性编码

• Flow Matching (ICLR 2023, Lipman et al.) — OT 视角简化扩散

• DSM ≡ Score Matching (Vincent 2011) — 去噪自编码 = 分数匹配

• Information Bottleneck (ISIT 2000, Tishby) — 压缩-相关性信息论框架

• MPPI (ICRA 2016/2017, Williams et al.) — Feynman-Kac 路径积分控制

• TD-MPC2 (ICLR 2024, Hansen et al.) — 317M 跨 80 任务模型预测控制

---

↑ 上级 · World Model