完整论文列表

本文收录 World Model、VLA、VLM 及无人机相关的核心论文 195 篇。 推荐等级：★★★ 必读 | ★★☆ 推荐 | ★☆☆ 选读

本文是元数据唯一权威：年份、标题、会议/期刊以本表为准。各专题文档与论文卡片只写结论并链回这里， 不重复维护年份；标题一律取 arXiv API 返回的完整标题 —— 系统名（如 NavFoM）是简称，不进标题位。

每一条写成 - **[会议'年] 名称** — *完整标题* · 推荐度 一行，下一行缩进挂徽章： arXiv 编号取红底 arXiv，项目页取蓝底 Website，代码库取黑底 GitHub。 有系统名的条目加粗位放系统名、斜体位放完整标题（如 TD-MPC/TD-MPC2）； 没有系统名的条目加粗位直接是完整标题、斜体位是作者（如 World Models）。 徽章只挂在清单层；各专题正文里的表格仍用纯文本链接，不挂徽章。

项目页与代码库收的是作者在摘要或 arXiv comment 里写出的那些，外加本仓库卡片里已核对可访问的那些； 两者都没有的条目只留 arXiv 一枚徽章（注意这不等于没有代码）。 按此口径，无人机三节 20 篇里有 12 篇只留 arXiv 徽章（60%），其余五节 61 篇里只有 19 篇（31%）。 这一项由此带了信息：空中方向的项目页与代码链接，比地面方向稀疏。 加上追前沿新增的条目，全表 195 篇里有 139 篇只留 arXiv 一枚徽章。

上面那组比例只对 2026-10 追前沿之前的 81 篇成立（分母 20 + 61 = 81）。此后的新增条目 项目页与代码库未逐条核查，与上面「只留 arXiv 一枚徽章」的含义不同，统计时不要混进来：

批次	条数	其中只有 arXiv 徽章	说明
波 1：1.5–1.7 三节与「二」的后 5 条	52	52	全部未核查
波 2：8.1–8.3 三节	62	58	4 条经 HTTP 核对后填入（BLIP-2、DINOv2、POPE、MME-RealWorld）
合计待核查	114	110	恢复信息量需逐条查过代码库

上表「4 条经 HTTP 核对」指只确认了 URL 返回 200，不等于该仓库与论文一一对应； 这一层核对在排版改造那一轮统一做。

---

一、世界模型 — 通用 (World Models, General)

1.1 经典基础

• [NeurIPS'18] World Models — Ha & Schmidhuber · ★★★
  

• [ICLR'20] Dream to Control — Dream to Control: Learning Behaviors by Latent Imagination · ★★★
  

• [ICLR'21] Mastering Atari with Discrete World Models — Hafner et al. · ★★★
  

• [Nature'25] Mastering Diverse Domains through World Models — Hafner et al. · ★★★
  Nature 2025 版标题为《Mastering diverse control tasks through world models》（DOI 10.1038/s41586-025-08744-2）。

• [ICLR'22/24] TD-MPC/TD-MPC2 — TD-MPC/TD-MPC2: Temporal Difference Learning for Model Predictive Control · ★★☆
  

1.2 Transformer 时代

• [ICLR'23] IRIS — Transformers are Sample-Efficient World Models (IRIS) · ★★★
  

• [NeurIPS'24] DIAMOND — Diffusion for World Modeling: Visual Details Matter in Atari (DIAMOND) · ★★★
  

• [arXiv'24.08] GameNGen — GameNGen: Diffusion Models are Real-Time Game Engines · ★★☆
  

• [NeurIPS'24] Learning Universal Policies via Text-Guided Video Generation — Du et al. · ★★☆
  

• [Report'24] Genie 2 — Genie 2: A Large-Scale Foundation World Model · ★★☆
  

1.3 大规模世界模型

• [arXiv'24.02] Genie — Genie: Generative Interactive Environments · ★★★
  

• [arXiv'23.10] UniSim — UniSim: Learning Interactive Real-World Simulators · ★★★
  

• [Report'25] Cosmos World Foundation Model Platform for Physical AI — NVIDIA · ★★★
  

• [Report'25] GAIA-2 — GAIA-2: Generative AI for Autonomous Driving · ★★☆
  

• [Report'25] World Labs — World Labs: Generating 3D Worlds · ★☆☆
  

1.4 世界模型理论与综述

• [IEEE T-ITS'24] World Models for Autonomous Driving — World Models for Autonomous Driving: An Initial Survey · ★★☆
  

• [arXiv'23.11] Diffusion Models for Reinforcement Learning: A Survey — Zhu et al. · ★☆☆
  

• [NeurIPS'22] Multi-Game Decision Transformers — Lee et al. · ★☆☆
  

1.5 联合嵌入预测与潜空间世界模型（JEPA 路线）

• [arXiv'23.01] Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture — Assran et al. · ★★★
  

• [arXiv'25.06] V-JEPA 2 — V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning · ★★★
  

• [arXiv'24.06] Cognitively Inspired Energy-Based World Models — Gladstone et al. · ★★☆
  

• [arXiv'26.06] Sensorimotor World Models — Sensorimotor World Models: Perception for Action via Inverse Dynamics · ★★☆
  

• [arXiv'26.07] ODEWorld — ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow · ★★☆
  

• [arXiv'26.08] Flow-JEPA — Flow-JEPA: Robust Latent Dynamics for JEPA World Models via Flow Matching · ★★☆
  

• [arXiv'26.09] MA-JEPA — MA-JEPA: Joint-Embedding World Models for Multi-Agent Reinforcement Learning · ★☆☆
  

• [arXiv'26.10] DeepJEPA — DeepJEPA: Scaling World Models from Within · ★★★
  

• [arXiv'26.10] CF-JEPA — CF-JEPA: Improving Robustness of JEPA World Models via Controllability Factorization · ★★☆
  

• [arXiv'26.10] Drive vs. Decay — Drive vs. Decay: On the Training Dynamics of Joint-Embedding Predictive Architectures · ★★☆
  

• [arXiv'26.10] TwinJEPA — TwinJEPA: Action-Preferred Predictive Representations for Goal-Conditioned Control · ★☆☆
  

• [arXiv'26.10] Keeping JEPA World Models Plannable When Little of the Frame Moves — Strohm et al. · ★★☆
  

• [arXiv'26.10] AVL-JEPA — AVL-JEPA: Preventing Causal Dynamics Information Collapse in Joint Embedding Predictive Architecture World Models · ★★☆
  

• [arXiv'26.10] Frozen in a Frame — Frozen in a Frame: The Velocity Blind Spot in JEPA World Models · ★★★
  

• [arXiv'26.10] H-JEPA — H-JEPA: End-to-End Learning of Hierarchical World Models for Visual Planning · ★★☆
  

• [arXiv'26.10] Preserving Unstable Modes Through Inverse Dynamics in JEPA World Models — Toso et al. · ★★☆
  

• [arXiv'26.10] Modeling Latent Disturbances for Robust Decision-Making in World Models — Seo et al. · ★☆☆
  

• [arXiv'26.10] DSReg — DSReg: Provably Recovering Individual World Latents without Reconstruction · ★★☆
  

• [arXiv'26.10] RoboJEPA — RoboJEPA: Scaling Robotic Latent World Models · ★★★
  

1.6 世界模型评测与诊断

• [arXiv'26.09] Same World, Different Knowledge — Same World, Different Knowledge: When Isolated Audits Misjudge World-Model Repairs · ★★★
  

• [arXiv'26.09] Beyond One-Step Accuracy — Beyond One-Step Accuracy: State-Affine Latent Transition for Reliable Visual Planning · ★★☆
  

• [arXiv'26.10] Learning Commute-Time-Preserving World Models for Planning — Hauri et al. · ★☆☆
  

• [arXiv'26.10] PROWBench — PROWBench: Do Video Models Render What the Program Specifies? · ★★☆
  

• [arXiv'26.10] Does Physics Live in the Activations? Localizing Physical Quantities in Video Diffusion Models — Kneifl et al. · ★★☆
  

• [arXiv'26.10] World Embedding Benchmark — Liu et al. · ★★☆
  

• [arXiv'26.10] When Low Prediction Error Misleads Planning: Diagnosing Representation, Dynamics, and Decision Failures in Latent World Models — Min et al. · ★★★
  

• [arXiv'26.10] World Models' Last Exam in Physics — Gao et al. · ★★☆
  

• [arXiv'26.10] World Models Dream of Success — World Models Dream of Success: Diagnosing and Repairing Failure Insensitivity in Robot World Models · ★★★
  

• [arXiv'26.10] Predicted Futures Are Not Enough — Predicted Futures Are Not Enough: Learning Executable Goals for Robot Manipulation · ★★☆
  

1.7 长时程世界模型与交互式生成

• [arXiv'26.08] GeniWorld — GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions · ★★☆
  

• [arXiv'26.08] ReWorld — ReWorld: An Interactive World Model with Long-Horizon Memory · ★★☆
  

• [arXiv'26.08] Matrix-Game 3.5 — Matrix-Game 3.5: Enhancing Real-Time Streaming Interactive World Models with Patch Memory · ★★☆
  

• [arXiv'26.09] SolarWM — SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models · ★☆☆
  

• [arXiv'26.09] The Past Frames the Future — The Past Frames the Future: Memory for Autoregressive Video Generation · ★★☆
  

• [arXiv'26.09] WorldPlay2 — WorldPlay2: Extending Real-Time Interactive World Models in Control and Horizon · ★☆☆
  

• [arXiv'26.09] Waypoint-1.5 — Waypoint-1.5: A Real-Time Video World Model for Consumer Hardware · ★★☆
  

• [arXiv'26.09] Beyond a Single Latent Space — Beyond a Single Latent Space: A Dual-Latent World Model for Long-Horizon Planning · ★★☆
  

• [arXiv'26.09] Honeycomb — Honeycomb: Constant-Size Scene Memory Representation for Video World Models · ★★★
  

• [arXiv'26.10] Oneira — Oneira: From Open-Ended Generation to Open-World Interaction in Video World Models · ★☆☆
  

• [arXiv'26.10] World Observer — World Observer: Joint Actor-Observer Generation for Persistent World Modeling · ★☆☆
  

• [arXiv'26.10] Spatial Memory Intelligence — Spatial Memory Intelligence: Endowing World Models with Understanding-Driven Long-Term Memory · ★★☆
  

• [arXiv'26.10] DeltaWorld — DeltaWorld: Physically Consistent Interactive World Simulators via Action-Conditioned Latent Increment Learning · ★★☆
  

• [arXiv'26.10] FLEX-WAM — FLEX-WAM: Flexible Block-Causal World-Action Models for Long-Horizon Imagination and Planning · ★★☆
  

• [arXiv'26.10] HLA-WM — HLA-WM: Hybrid Linear Attention for Long-Horizon Video World Models · ★★☆
  

• [arXiv'26.10] SimForcing — SimForcing: Distilling Simulation Motion Priors into Real-Domain Robot World Models · ★★☆
  

• [arXiv'26.10] Tracking Is Not Permanence — Tracking Is Not Permanence: What Video World Models Keep of a Hidden Object · ★★★
  

• [arXiv'26.10] Parallel Predictive World Models for Accurate and Efficient Long-Horizon Planning — Feng et al. · ★★☆
  

1.5–1.7 三节是 2026-10 追前沿时新增的（对应 docs/02-世界模型专题/07–09 三篇）。这三节的项目页与代码库没有逐条核查， 只留 arXiv 一枚徽章表示「未核查」，不等于「没有代码」——与其余各节只留 arXiv 徽章的含义不同，统计时不要混。 三节里的 arXiv 号取自 arXiv API；录用去向在 API 返回里拿不到，标签一律写 [arXiv'YY.MM]，不臆测。

---

二、世界模型 — 无人机 (World Models, UAV/Drone)

• [IEEE RA-L'24] Learning to Fly in Seconds — Eschmann et al. · ★★☆
  

• [ICRA'25] Dream to Fly — Dream to Fly: Model-Based Reinforcement Learning for Vision-Based Drone Flight · ★★★
  

• [arXiv'25.12] Aerial World Model for Long-horizon Visual Generation and Navigation in 3D Space — Zhang et al. · ★★★
  

• [arXiv'25.09] FlightDiffusion — FlightDiffusion: Revolutionising Autonomous Drone Training with Diffusion Models Generating FPV Video · ★★☆
  

• [arXiv'26.04] MotionScape — MotionScape: A Motion-Stratified UAV Video Benchmark for World Modeling and Future Video Generation · ★★☆
  

• [arXiv'24.08] AeroVerse — AeroVerse: UAV-Agent Benchmark Suite for Simulating, Pre-training, Finetuning, and Evaluating Aerospace Embodied World Models · ★★☆
  

• [arXiv'26.05] Aero-World — Aero-World: Action-Conditioned Aerial Video Generation from Inertial Controls · ★★★
  

• [arXiv'26.05] WorldVLN — WorldVLN: Autoregressive World Action Model for Aerial Vision-Language Navigation · ★★☆
  

• [arXiv'26.06] WorldFly — WorldFly: A World-Model-Based Vision-Language-Action Model for UAV Navigation · ★★☆
  

• [arXiv'26.09] Skytopia — Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models · ★★★
  

• [arXiv'26.09] ForeFly — ForeFly: A Dual-Horizon World Action Model for Aerial Vision-Language Navigation · ★★★
  

注：本节仅保留已验证 arXiv ID 的论文。更多无人机世界模型论文请参考 无人机世界模型综述。

后 5 条（Aero-World / WorldVLN / WorldFly / Skytopia / ForeFly）是 2026-10 追前沿时新增的，其中 2605.19728 / 2605.15964 / 2606.06147 三条落在 tools/watch.py 的 120 天窗口起点（2026-06-10）之前几周，是换用不限窗口的查询复核时才捞出来的—— 复核证据记在 tools/watchlists.py 的 VERIFIED_ZEROS 里。窗口起点会切掉一整批同主线的论文，写「近期空白」之前必须先看不限窗口的命中。

---

三、VLA — 通用 (Vision-Language-Action, General)

3.1 奠基工作

• [ICML'23] PaLM-E — PaLM-E: An Embodied Multimodal Language Model · ★★★
  

• [CoRL'23] RT-2 — RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control · ★★★
  

• [RSS'23] RT-1 — RT-1: Robotics Transformer for Real-World Control · ★★☆
  

• [CoRL'22] SayCan — SayCan: Do As I Can, Not As I Say · ★★☆
  

3.2 开源时代

• [arXiv'24.06] OpenVLA — OpenVLA: An Open-Source Vision-Language-Action Model · ★★★
  

• [arXiv'24.05] Octo — Octo: An Open-Source Generalist Robot Policy · ★★★
  

• [ICRA'24] Open X-Embodiment — Open X-Embodiment: Robotic Learning Datasets and RT-X Models · ★★★
  

• [arXiv'24.12] RoboVLMs — What Matters in Building Vision-Language-Action Models for Generalist Robots (RoboVLMs) · ★★☆
  

• [arXiv'24.01] SpatialVLM — SpatialVLM: Endowing Vision-Language Models with Spatial Reasoning · ★☆☆
  

3.3 流匹配与扩散 VLA

• [arXiv'24.10] π₀ — π₀: A Vision-Language-Action Flow Model for General Robot Control · ★★★
  

• [arXiv'25.04] π₀.₅ — π₀.₅: a Vision-Language-Action Model with Open-World Generalization · ★★★
  

• [RSS'24] 3D Diffusion Policy (DP3) — 3D Diffusion Policy (DP3): Generalizable Visuomotor Policy · ★★☆
  

• [RSS'23] Diffusion Policy — Diffusion Policy: Visuomotor Policy Learning via Action Diffusion · ★★☆
  

3.4 VLA 理论与综述

• [arXiv'24.05] A Survey on Vision-Language-Action Models for Embodied AI — Ma et al. · ★★★
  

• [arXiv'23.03] Foundation Models for Decision Making — Foundation Models for Decision Making: Problems, Methods, and Opportunities · ★★☆
  

• [arXiv'25.01] FAST — FAST: Efficient Action Tokenization for Vision-Language-Action Models · ★★☆
  

---

四、VLA — 无人机 (Vision-Language-Action, UAV/Drone)

• [arXiv'25.12] VLA-AN — VLA-AN: An Efficient and Onboard Vision-Language-Action Framework for Aerial Navigation in Complex Environments · ★★★
  

• [arXiv'25.03] CognitiveDrone — CognitiveDrone: A VLA Model and Evaluation Benchmark for Real-Time Cognitive Task Solving and Reasoning in UAVs · ★★★
  

• [arXiv'26.04] UAV-Track VLA — UAV-Track VLA: Embodied Aerial Tracking via Vision-Language-Action Models · ★★★
  

• [arXiv'25.05] UAV-Flow Colosseo — UAV-Flow Colosseo: A Real-World Benchmark for Flying-on-a-Word UAV Imitation Learning · ★★☆
  

• [arXiv'26.02] VLN-Pilot — VLN-Pilot: Large Vision-Language Model as an Autonomous Indoor Drone Operator · ★★☆
  

• [ICLR'26] AutoFly — AutoFly: Vision-Language-Action Model for UAV Autonomous Navigation in the Wild · ★★☆
  

---

五、VLM — 遥感 (Vision-Language Models, Remote Sensing)

• [CVPR'24] GeoChat — GeoChat: Grounded Large Vision-Language Model for Remote Sensing · ★★★
  

• [arXiv'23.07] RSGPT — RSGPT: A Remote Sensing Vision Language Model and Benchmark · ★★★
  

• [ECCV'24] LHRS-Bot — LHRS-Bot: Empowering Remote Sensing with VGI-Enhanced Large Multimodal Language Model · ★★☆
  

• [arXiv'24.09] ChangeChat — ChangeChat: An Interactive Model for Remote Sensing Change Analysis via Multimodal Instruction Tuning · ★★☆
  

• [arXiv'24.01] EarthGPT — EarthGPT: A Universal Multi-modal Large Language Model for Multi-sensor Image Comprehension in Remote Sensing Domain · ★★☆
  

• [arXiv'24.01] SkyEyeGPT — SkyEyeGPT: Unifying Remote Sensing Vision-Language Tasks via Instruction Tuning with Large Language Model · ★☆☆
  

• [arXiv'24.02] ChatEarthNet — ChatEarthNet: A Global-Scale Image-Text Dataset Empowering Vision-Language Geo-Foundation Models · ★☆☆
  

• [TGRS'24] RemoteCLIP — RemoteCLIP: A Vision Language Foundation Model for Remote Sensing · ★★☆
  

---

六、VLM — 无人机 (Vision-Language Models, UAV/Drone)

• [arXiv'26.01] DVGBench — DVGBench: Implicit-to-Explicit Visual Grounding Benchmark in UAV Imagery with Large Vision-Language Models · ★★★
  

• [arXiv'26.03] UAVBench and UAVIT-1M — UAVBench and UAVIT-1M: Benchmarking and Enhancing MLLMs for Low-Altitude UAV Vision-Language Understanding · ★★☆
  

• [arXiv'26.04] Can VLMs Think from the Sky? Unifying UAV Reasoning and Generation — Sun et al. · ★★☆
  

• [arXiv'25.12] CoDrone — CoDrone: Autonomous Drone Navigation Assisted by Edge and Cloud Foundation Models · ★★☆
  

• [arXiv'25.05] FM-Planner — FM-Planner: Foundation Model Guided Path Planning for Autonomous Drone Navigation · ★★☆
  

• [arXiv'25.09] Embodied Navigation Foundation Model — Zhang et al. · ★★☆
  

• [ACL'25] CityNavAgent — CityNavAgent: Aerial Vision-and-Language Navigation with Hierarchical Semantic Planning and Global Memory · ★★☆
  

• [arXiv'25.03] UAV-VLRR — UAV-VLRR: Vision-Language Informed NMPC for Rapid Response in UAV Search and Rescue · ★★☆
  

注：更多无人机 VLM 论文请参考 无人机场景理解 和 LLM驱动的无人机Agent。

---

七、基准与数据集 (Benchmarks & Datasets)

• [ICRA'24] Open X-Embodiment — Open X-Embodiment: Robotic Learning Datasets and RT-X Models · ★★★
  

• [arXiv'23.06] LIBERO — LIBERO: Benchmarking Knowledge Transfer in Lifelong Robot Learning · ★★☆
  

• [arXiv'24.05] SIMPLER — Evaluating Real-World Robot Manipulation Policies in Simulation (SIMPLER) · ★★☆
  

• [ICLR'23] ManiSkill2 — ManiSkill2: A Unified Benchmark for Generalizable Manipulation · ★★☆
  

• [ICLR'24] Habitat 3.0 — Habitat 3.0: A Co-Habitat for Humans, Avatars, and Robots · ★★☆
  

• [ISER'18] AirSim — AirSim: High-Fidelity Visual and Physical Simulation · ★★★
  

• [Various'20-24] DroneCrowd / DroneVehicle / VisDrone Benchmarks — Various · ★★☆
  

• [arXiv'21.08] TS4Net — TS4Net: Two-Stage Sample Selective Strategy for Rotating Object Detection（提出 UAV-ROD 无人机数据集） · ★☆☆
  

• [arXiv'25.11] Is your VLM Sky-Ready? A Comprehensive Spatial Intelligence Benchmark for UAV Navigation · ★☆☆
  

• [arXiv'26.01] AIR-VLA — AIR-VLA: Vision-Language-Action Systems for Aerial Manipulation · ★★☆
  

---

八、多模态基础模型与通用 VLM 方法 (Foundation Models & General VLM)

本节按通用 VLM 的方法主干分三组，与 docs/04-VLM专题/ 的 05（架构与编码器）、06（指令微调）、07（评测与幻觉）三篇一一对应。### 8.x 子编号是本次新增的，未改动「五」至「十」的节号。

8.1 架构与视觉编码器（→ 04-VLM专题/05）

• [ICML'21] CLIP — CLIP: Learning Transferable Visual Models from NLP · ★★★
  

• [CVPR'21] LiT — LiT: Zero-Shot Transfer with Locked-image text Tuning · ★☆☆
  

• [NeurIPS'22] Flamingo — Flamingo: a Visual Language Model for Few-Shot Learning · ★★★
  

• [CVPR'22] Scaling Language-Image Pre-training via Masking — Li et al. · ★☆☆
  

• [arXiv'23.01] BLIP-2 — BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models · ★★★
  

• [arXiv'23.03] EVA-CLIP — EVA-CLIP: Improved Training Techniques for CLIP at Scale · ★☆☆
  

• [ICCV'23] SigLIP — SigLIP: Sigmoid Loss for Language-Image Pre-Training · ★★★
  

• [arXiv'25.02] SigLIP 2 — SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features · ★★☆
  

• [NeurIPS'23] DataComp — DataComp: In search of the next generation of multimodal datasets · ★☆☆
  

• [arXiv'23.04] DINOv2 — DINOv2: Learning Robust Visual Features without Supervision · ★★☆
  

• [arXiv'23.07] Patch n' Pack — Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution · ★★☆
  

• [arXiv'23.08] Qwen-VL — Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond · ★★☆
  

• [arXiv'24.09] Qwen2-VL — Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution · ★★★
  

• [arXiv'25.02] Qwen2.5-VL Technical Report — Qwen Team · ★★★
  

• [CVPR'24] InternVL — InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks · ★★☆
  

• [arXiv'24.12] Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling — Chen et al. · ★★☆
  

• [CVPR'23] Honeybee — Honeybee: Locality-enhanced Projector for Multimodal LLM · ★★☆
  

• [ICLR'22] Token Merging — Token Merging: Your ViT But Faster · ★★☆
  

• [arXiv'24.02] MobileVLM V2 — MobileVLM V2: Faster and Stronger Baseline for Vision Language Model · ★☆☆
  

• [ICCV'24] LLaVA-PruMerge — LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models · ★☆☆
  

• [arXiv'24.05] Matryoshka Multimodal Models — Cai et al. · ★★☆
  

• [arXiv'24.05] Ovis — Ovis: Structural Embedding Alignment for Multimodal Large Language Model · ★☆☆
  

• [arXiv'24.07] TokenPacker — TokenPacker: Efficient Visual Projector for Multimodal LLM · ★☆☆
  

• [ICML'24] Prismatic VLMs — Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models · ★★☆
  

• [arXiv'24.03] MM1 — MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training · ★★★
  

• [arXiv'24.05] What matters when building vision-language models? — Laurençon et al. · ★★☆
  

• [arXiv'24.06] Cambrian-1 — Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs · ★☆☆
  

• [arXiv'24.08] LLaVA-OneVision — LLaVA-OneVision: Easy Visual Task Transfer · ★★☆
  

• [Report'23] GPT-4V(ision) System Card — OpenAI · ★★☆
  

8.2 指令微调与对齐（→ 04-VLM专题/06）

• [NeurIPS'23] LLaVA — LLaVA: Visual Instruction Tuning · ★★★
  

• [arXiv'23.10] LLaVA-1.5 — LLaVA-1.5: Improved Baselines with Visual Instruction Tuning · ★★★
  

• [arXiv'23.04] MiniGPT-4 — MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models · ★★☆
  

• [arXiv'23.05] InstructBLIP — InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning · ★★☆
  

• [arXiv'23.04] mPLUG-Owl — mPLUG-Owl: Modularization Empowers Large Language Models with Multimodality · ★☆☆
  

• [TPAMI'23] Otter — Otter: A Multi-Modal Model with In-Context Instruction Tuning · ★☆☆
  

• [arXiv'23.11] SPHINX — SPHINX: The Joint Mixing of Weights, Tasks, and Visual Embeddings for Multi-modal Large Language Models · ★☆☆
  

• [arXiv'23.09] InternLM-XComposer — InternLM-XComposer: A Vision-Language Large Model for Advanced Text-image Comprehension and Composition · ★☆☆
  

• [arXiv'23.07] SVIT — SVIT: Scaling up Visual Instruction Tuning · ★★☆
  

• [arXiv'23.11] ShareGPT4V — ShareGPT4V: Improving Large Multi-Modal Models with Better Captions · ★★☆
  

• [arXiv'24.02] ALLaVA — ALLaVA: Harnessing GPT4V-Synthesized Data for Lite Vision-Language Models · ★☆☆
  

• [arXiv'24.02] TinyLLaVA — TinyLLaVA: A Framework of Small-scale Large Multimodal Models · ★☆☆
  

• [ICLR'21] LoRA — LoRA: Low-Rank Adaptation of Large Language Models · ★★★
  

• [NeurIPS'23] QLoRA — QLoRA: Efficient Finetuning of Quantized LLMs · ★★☆
  

• [ICLR'23] LLaMA-Adapter — LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention · ★★☆
  

• [NeurIPS'23] Direct Preference Optimization — Direct Preference Optimization: Your Language Model is Secretly a Reward Model · ★★★
  

• [arXiv'23.09] Aligning Large Multimodal Models with Factually Augmented RLHF — Sun et al. · ★☆☆
  

• [CVPR'23] RLHF-V — RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback · ★★☆
  

• [National Science Review'23] A Survey on Multimodal Large Language Models — Yin et al. · ★★☆
  

• [arXiv'23.09] Multimodal Foundation Models — Multimodal Foundation Models: From Specialists to General-Purpose Assistants · ★★☆
  

8.3 评测与幻觉（→ 04-VLM专题/07）

• [arXiv'23.09] The Dawn of LMMs — The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision) · ★★☆
  

• [CVPR'23] MMMU — MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI · ★★★
  

• [ACL'24] MMMU-Pro — MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark · ★★☆
  

• [ECCV'23] MMBench — MMBench: Is Your Multi-modal Model an All-around Player? · ★★★
  

• [NeurIPS'23] MME — MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models · ★★☆
  

• [ICML'23] MM-Vet — MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities · ★★☆
  

• [ICLR'23] MathVista — MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts · ★★☆
  

• [arXiv'24.03] Are We on the Right Way for Evaluating Large Vision-Language Models? — Chen et al. · ★★☆
  

• [CVPR'23] MVBench — MVBench: A Comprehensive Multi-modal Video Understanding Benchmark · ★☆☆
  

• [ICLR'24] MME-RealWorld — MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans? · ★★☆
  

• [ICML'24] MLLM-as-a-Judge — MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark · ★☆☆
  

• [CVPR'24] LLaVA-Critic — LLaVA-Critic: Learning to Evaluate Multimodal Models · ★☆☆
  

• [EMNLP'23] What's "up" with vision-language models? Investigating their struggle with spatial reasoning — Kamath et al. · ★★★
  

• [ECCV'24] BLINK — BLINK: Multimodal Large Language Models Can See but Not Perceive · ★★☆
  

• [arXiv'24.01] Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs — Tong et al. · ★★☆
  

• [EMNLP'23] Evaluating Object Hallucination in Large Vision-Language Models — Li et al. · ★★★
  

• [CVPR'23] HallusionBench — HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models · ★★★
  

• [arXiv'23.11] AMBER — AMBER: An LLM-free Multi-dimensional Benchmark for MLLMs Hallucination Evaluation · ★★☆
  

• [arXiv'24.02] A Survey on Hallucination in Large Vision-Language Models — Liu et al. · ★★☆
  

• [arXiv'23.11] Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding — Leng et al. · ★★☆
  

• [CVPR'23] OPERA — OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation · ★★☆
  

• [SCI CHINA Inf. Sci.'23] Woodpecker — Woodpecker: Hallucination Correction for Multimodal Large Language Models · ★☆☆
  

---

九、统计摘要

pie title 论文分布
    "世界模型(通用)" : 65
    "世界模型(无人机)" : 11
    "VLA(通用)" : 16
    "VLA(无人机)" : 6
    "VLM(遥感)" : 8
    "VLM(无人机)" : 8
    "基准/数据集" : 10
    "多模态与通用VLM" : 71

类别	论文数	必读(★★★)	推荐(★★☆)	选读(★☆☆)
世界模型 — 通用	65	19	35	11
世界模型 — 无人机	11	5	6	0
VLA — 通用	16	8	7	1
VLA — 无人机	6	3	3	0
VLM — 遥感	8	2	4	2
VLM — 无人机	8	1	7	0
基准与数据集	10	2	6	2
多模态基础模型与通用 VLM	71	16	35	20
合计	195	56	103	36

「世界模型 — 通用 65 篇」里含 2026-10 追前沿新增的 1.5 JEPA 路线（19 篇）、1.6 评测与诊断（10 篇）、 1.7 长时程与交互式生成（18 篇）三节，共 47 篇；「世界模型 — 无人机」的 11 篇里有 5 篇同批新增。

「多模态基础模型与通用 VLM 71 篇」是按 8.1 架构与视觉编码器（29 篇）、8.2 指令微调与对齐（20 篇）、 8.3 评测与幻觉（22 篇）三组对应 04-VLM专题 的 05–07 三篇；原先的「多模态基础模型」9 篇 （CLIP / SigLIP / LLaVA / LLaVA-1.5 / Qwen-VL / Qwen2.5-VL / InternVL / Cambrian-1 / GPT-4V）全部并入这三组。 没有删除，原先的条目序号已随条目式改写弃用。8.3 一组里的评测基准与「七、基准与数据集」不重复：那一节收的是机器人/无人机的 数据与评测协议，这一组收的是通用 VLM 的基准与幻觉诊断。

交叉列出现在是有意的。 Open X-Embodiment（arXiv:2310.08864）同时出现在「三、VLA — 通用」与「七、基准与数据集」， 因为它既是一个 VLA 训练语料库、也是一个评测基准。凡同时是数据集与基准的工作都可能出现两次，不是重复录入。

AIR-VLA（arXiv:2601.21602）名字里带 VLA，但产品是数据与评测协议，不是模型，所以收在「七、基准与数据集」， 不在「四、VLA — 无人机」—— 分类按产出物走，不按标题走。

---

十、引用格式

如需 BibTeX 格式，推荐使用 Semantic Scholar 或 Google Scholar 导出。以下为示例：

@article{brohan2023rt2,
  title={RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control},
  author={Brohan, Anthony and others},
  journal={arXiv preprint arXiv:2307.15818},
  year={2023}
}

@article{hafner2023dreamerv3,
  title={Mastering Diverse Domains through World Models},
  author={Hafner, Danijar and others},
  journal={Journal of Machine Learning Research},
  year={2023}
}

@article{black2024pi0,
  title={pi{\_}0: A Vision-Language-Action Flow Model for General Robot Control},
  author={Black, Kevin and others},
  journal={arXiv preprint arXiv:2410.24164},
  year={2024}
}

---

推荐等级（★★★ 必读 / ★★☆ 推荐 / ★☆☆ 选读）根据论文的影响力、创新性和与无人机领域的相关性综合评定。会议/期刊信息以正式录用版本为准，部分预印本可能后续变更。

本文件为 UAV-WM-VLA-Learning 项目的一部分，最后更新：2026-10-09。