Awesome Agentic World Modeling

August 22, 2026 · View on GitHub

Awesome Agentic World Modeling

This repository accompanies the position paper "Quo Vadis, World Modeling? Towards Interactive World Proxies for Continually Improving Agents," and curates the works it reviews. The survey argues for a shift from a World Model that passively predicts the next physical state toward an Agent-Centric World Proxy: an environment-grounded interface that returns the information transition an agent needs (a future state, rendered view, execution result, retrieved memory or skill, or verdict on a plan) to plan, learn, and continually improve.

The taxonomy has two orthogonal axes:

Six proxy functions (what feedback the proxy returns): Dynamics, Spatial, Execution, Memory / Experience, Skill, and Reward / Verification.
Three empowerment levels (how the proxy improves the agent): L1 inference-time guidance, L2 training-time optimization, and L3 Agent-Proxy co-evolution.

Call for Contributions

Agentic world modeling moves fast, and the boundary between a passive world model and an agent-facing world proxy is still being drawn. This list aims to be a living, community-maintained map of that space, and contributions of new work or corrections are very welcome.

To contribute, open an issue or a pull request:

Where: add each entry to the section matching its proxy function (or the closest one), and keep every table in chronological order.
Format: follow the existing row layout: a short model name in Model (or - if none), the title with an arXiv badge in Paper, and the abbreviated venue with year in Venue.
Verify: check the title, authors, arXiv ID, and venue against the primary source before adding, and prefer the formally published version whenever one exists.
Fix: corrections to venues, broken links, or mis-categorized entries are equally welcome; flag them in an issue or fix them directly in a pull request.

Background
- Classical World Models & Model-Based RL
- Continual Improvement & Interaction Cost
The Six Proxy Functions
Empowerment Levels (L1 / L2 / L3)
Evaluation, Fidelity & Safety
Surveys & Related Work
Additional References
Citation

Background

Classical World Models & Model-Based RL

Foundations: state-transition world models and model-based RL that predict the next physical state given a state and an action.

Model	Paper	Venue

`-`	Dyna, an integrated architecture for learning, planning, and reacting	SIGART 1991
`-`	Recurrent World Models Facilitate Policy Evolution	NeurIPS 2018
`-`	When to Trust Your Model: Model-Based Policy Optimization	NeurIPS 2019
`-`	Dream to Control: Learning Behaviors by Latent Imagination	ICLR 2020
`-`	Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model	Nature 2020
`-`	A Path Towards Autonomous Machine Intelligence	OpenReview preprint 2022
`-`	Model-Based Reinforcement Learning: A Survey	FTML 2023
`-`	Mastering diverse control tasks through world models	Nature 2025

Continual Improvement & Interaction Cost

Why real-environment interaction alone cannot carry continual improvement: cost, safety, latency, and the need for scalable feedback.

Model	Paper	Venue

`MuJoCo`	MuJoCo: A Physics Engine for Model-Based Control	IROS 2012
`-`	Isaac Gym: High Performance GPU-Based Physics Simulation for Robot Learning	NeurIPS 2021
`DayDreamer`	DayDreamer: World Models for Physical Robot Learning	CoRL 2023
`CodeIt`	CodeIt: Self-Improving Language Models with Prioritized Hindsight Replay	ICML 2024
`Voyager`	Voyager: An Open-Ended Embodied Agent with Large Language Models	TMLR 2024
`OSWorld`	OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments	NeurIPS 2024
`WebArena`	WebArena: A Realistic Web Environment for Building Autonomous Agents	ICLR 2024
`CASCADE`	CASCADE: Cumulative Agentic Skill Creation through Autonomous Development and Evolution	arXiv 2025
`-`	Continual Learning and Catastrophic Forgetting	Learning and Memory: A Comprehensive Reference 2025
`-`	Agent Learning via Early Experience	ICML 2026

The Six Proxy Functions

Each proxy function answers a different question for the agent, but all return agent-usable feedback.

Dynamics Proxy

How would the environment change if the agent executed a given action? Video, driving, game, and model-based dynamics predictors.

Model	Paper	Venue

`-`	Learning Latent Dynamics for Planning from Pixels	ICML 2019
`FitVid`	FitVid: Overfitting in Pixel-Level Video Prediction	arXiv 2021
`VideoGPT`	VideoGPT: Video Generation using VQ-VAE and Transformers	arXiv 2021
`-`	Video Diffusion Models	NeurIPS 2022
`-`	Diffusion Models for Video Prediction and Infilling	TMLR 2022
`MCVD`	MCVD: Masked Conditional Video Diffusion for Prediction, Generation, and Interpolation	NeurIPS 2022
`GAIA-1`	GAIA-1: A Generative World Model for Autonomous Driving	arXiv 2023
`-`	Transformers Are Sample-Efficient World Models	ICLR 2023
`STORM`	STORM: Efficient Stochastic Transformer based World Models for Reinforcement Learning	NeurIPS 2023
`-`	Diffusion for World Modeling: Visual Details Matter in Atari	NeurIPS 2024
`-`	Video generation models as world simulators	2024
`Genie`	Genie: Generative Interactive Environments	ICML 2024
`Vista`	Vista: A Generalizable Driving World Model with High Fidelity and Versatile Controllability	NeurIPS 2024
`TD-MPC2`	TD-MPC2: Scalable, Robust World Models for Continuous Control	ICLR 2024
`RoboCasa`	RoboCasa: Large-Scale Simulation of Everyday Tasks for Generalist Robots	RSS 2024
`DriveDreamer`	DriveDreamer: Towards Real-World-Driven World Models for Autonomous Driving	ECCV 2024
`iVideoGPT`	iVideoGPT: Interactive VideoGPTs are Scalable World Models	NeurIPS 2024
`-`	Learning Interactive Real-World Simulators	ICLR 2024
`DynamicCity`	DynamicCity: Large-Scale 4D Occupancy Generation from Dynamic Scenes	ICLR 2025
`GameGen-X`	GameGen-X: Interactive Open-world Game Video Generation	ICLR 2025
`-`	Genie 3: A New Frontier for World Models	2025
`-`	Towards Unified World Models for Visual Navigation via Memory-Augmented Planning and Foresight	arXiv 2025
`-`	World Modelling Improves Language Model Agents	arXiv 2025
`MineWorld`	MineWorld: a Real-Time and Open-Source Interactive World Model on Minecraft	arXiv 2025
`MobileWorldBench`	MobileWorldBench: Towards Semantic World Modeling For Mobile Agents	arXiv 2025
`-`	World Model on Million-Length Video And Language With Blockwise RingAttention	ICLR 2025
`Vision`	Vision-Centric 4D Occupancy Forecasting and Planning via Implicit Residual World Models	arXiv 2025
`-`	V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning	arXiv 2025
`-`	Diffusion Models Are Real-Time Game Engines	ICLR 2025
`Marble`	Marble: A Multimodal World Model	2025
`-`	Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving	AAAI 2025
`Matrix-Game`	Matrix-Game: Interactive World Foundation Model	arXiv 2025
`Language`	Language-Conditioned World Modeling for Visual Navigation	arXiv 2026
`-`	Advancing Open-source World Models	arXiv 2026
`NavThinker`	NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation	arXiv 2026
`LiDARCrafter`	LiDARCrafter: Dynamic 4D World Modeling from LiDAR Sequences	AAAI 2026
`Self`	Self-Execution Simulation Improves Coding Models	arXiv 2026
`-`	Is Your Driving World Model an All-Around Player?	CVPR 2026
`U4D`	U4D: Uncertainty-Aware 4D World Modeling from LiDAR Sequences	CVPR 2026
`AD-R1`	AD-R1: Closed-Loop Reinforcement Learning for End-to-End Autonomous Driving with Impartial World Models	CVPR 2026
`SPIRAL`	SPIRAL: Self-Evolving Action-Conditioned Video Generation via Reflective Planning Agents	arXiv 2026
`FlowWAM`	FlowWAM: Optical Flow as a Unified Action Representation for World Action Models	arXiv 2026
`LeapBot`	LeapBot-WA: World-Anchor Action Models via Predictive Latent Alignments	arXiv 2026
`Enfold`	Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control	arXiv 2026
`-`	World Action Planner: Generalizable Decision-Making with Action-Conditioned World Models	arXiv 2026
`Auto`	Auto-JEPA: A Latent World Model of Continuous Intent for End-to-End Autonomous Driving	arXiv 2026
`Faster`	Faster-WAM: Do World Action Models Need Deep Action Modules?	arXiv 2026
`DF$^3$`	DF $^3$ : World Modeling via Decoder-Free Feature Forecasting in Autonomous Navigation	arXiv 2026
`SG`	SG-WAM: Self-Guided World Modeling in Geometry-Aware Policy Space	arXiv 2026
`LAWM`	LAWM-3D: Learning 3D-Aware Latent Actions from Human Videos for Generalizable Robot World Models	arXiv 2026
`XEWorld`	XEWorld: Can Action-Conditioned World Models Generalize to Unseen Robot Embodiments?	arXiv 2026
`Adaptive`	Adaptive-WAM: Quality-Guided Early-Exit Planning from Intermediate Video-Diffusion Features	arXiv 2026
`GWM`	GWM-VLA: Geometry-Aware Latent World Modeling for Vision-Language-Action Learning	arXiv 2026
`4D`	4D-WAM: Infusing Spatiotemporal Awareness into World Action Models through Trajectory Fields	arXiv 2026
`WA`	WA-SpecDec: World-Aware Speculative Decoding for Vision-Language-Action Models	arXiv 2026
`JEPA`	JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling	arXiv 2026
`-`	World Tokens: Enhancing Embodied Policies with Training-Time World Modeling	arXiv 2026
`SLIM`	SLIM-0.5B: Learning Action-Grounded Predictive Latents for Robot Manipulation	arXiv 2026
`Energy`	Energy-Structured Latent World Models with Neural Time Fields for Physically Constistent Open-World Motion Planning	arXiv 2026
`Dreamer`	Dreamer-SAC: Off-Policy Learning in Latent World Models for Sample-Efficient Autonomous Driving	arXiv 2026
`BrainWAM`	BrainWAM: Action-Space Coordination of Semantic Priors and Predictive Dynamics for Autonomous Driving	arXiv 2026
`DreamX`	DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation	arXiv 2026
`ForgeWM`	ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models	arXiv 2026
`-`	Q-Learning With World Models	arXiv 2026
`Hydra`	Hydra-0: Action Flow for Generalist World Modeling and Control	arXiv 2026
`GigaBrain`	GigaBrain-WBC-0.5: A Behavior World Model for Robust Whole-Body Control with Environment Interaction	arXiv 2026
`-`	Reinforced Planning with Latent World Models	arXiv 2026
`DA`	DA-WAM: Decision-Aligned Future Latents for Driving World Models	arXiv 2026
`-`	Temporally Centered SIGReg Improves Multi-Task LeWorldModel Learning: From Analysis to Method	arXiv 2026
`-`	Overcoming Statistical Bias in Action-Controllable World Models	arXiv 2026
`-`	Mitigating Compounding Error via Video Representation Regularization	arXiv 2026
`ODEWorld`	ODEWorld: A Continuous Predictive Architecture via Physical-Time Flow	arXiv 2026
`GeniWorld`	GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions	arXiv 2026
`ContactFlow`	ContactFlow: A video action conditioning that transfers across embodiments	arXiv 2026
`EmbodiedVAE`	EmbodiedVAE: Disentangled Video VAE for Efficient and Controllable Embodied Manipulation	arXiv 2026
`-`	$τ_0$ -VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation	arXiv 2026
`Depth`	Depth-Regularized JEPA World Models Learn More Transferable Representations from Real Outdoor Robot Data	arXiv 2026
`-`	Is Forward Prediction Enough? Physical State Grounding for JEPA World Models	arXiv 2026
`Xiaomi`	Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model	arXiv 2026
`MiniWorld`	MiniWorld: Democratizing the Training of Video World Models from Scratch	arXiv 2026
`BWM`	BWM: A Low-Cost High-Fidelity World Simulator for Robot Learning	arXiv 2026
`S2`	S2-HWM: Sparse Event-Structured Hierarchical World Model for Long-Horizon Surgical Robot Manipulation	arXiv 2026
`-`	Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers	arXiv 2026
`Population`	Population-Scalable Multi-Agent World Modeling	arXiv 2026
`DreamQAS`	DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search	arXiv 2026
`FedWorld`	FedWorld: Scope-Aware Federation of Agent World Models	arXiv 2026
`-`	Offline Multi-Agent Reinforcement Learning with a Physics-Informed World Model for Cooperative Mixed Traffic Control	arXiv 2026
`DreamTrajectory`	DreamTrajectory: Trajectory-Guided Action Generation with World Model Alignment for Mobile Manipulation	arXiv 2026
`EndoWAM`	EndoWAM: A Grounded World-Action Model for Generalizable Endoscopic Navigation	arXiv 2026
`-`	Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning	arXiv 2026

Spatial Proxy

What would the agent observe from another viewpoint or position? NeRF/Gaussian rendering, 3D/4D reconstruction and generation, and visual imagination for navigation.

Model	Paper	Venue

`NeRF`	NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis	ECCV 2020
`Mip-NeRF`	Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields	ICCV 2021
`PathDreamer`	PathDreamer: A World Model for Indoor Navigation	ICCV 2021
`Mip`	Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields	CVPR 2022
`SceneDreamer`	SceneDreamer: Unbounded 3D Scene Generation from 2D Image Collections	TPAMI 2023
`SceneScape`	SceneScape: Text-Driven Consistent Scene Generation	NeurIPS 2023
`Text2Room`	Text2Room: Extracting Textured 3D Meshes from 2D Text-to-Image Models	ICCV 2023
`VoxPoser`	VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models	CoRL 2023
`-`	3D Gaussian Splatting for Real-Time Radiance Field Rendering	TOG 2023
`Nerfstudio`	Nerfstudio: A Modular Framework for Neural Radiance Field Development	SIGGRAPH Asia 2023
`-`	A Survey on 3D Gaussian Splatting	CSUR 2024
`-`	3D Gaussian as a New Era: A Survey	TVCG 2024
`-`	Grounding Image Matching in 3D with MASt3R	ECCV 2024
`DreamScene`	DreamScene: 3D Gaussian-based Text-to-3D Scene Generation via Formation Pattern Sampling	ECCV 2024
`Scaffold-GS`	Scaffold-GS: Structured 3D Gaussians for View-Adaptive Rendering	CVPR 2024
`Dreamforge`	Dreamforge: Motion-aware autoregressive video generation for multi-view driving scenes	arXiv 2024
`DriveWorld`	DriveWorld: 4D Pre-trained Scene Understanding via World Models for Autonomous Driving	CVPR 2024
`DimensionX`	DimensionX: Create Any 3D and 4D Scenes from a Single Image with Controllable Video Diffusion	arXiv 2024
`ManiSkill3`	ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI	arXiv 2024
`DUSt3R`	DUSt3R: Geometric 3D Vision Made Easy	CVPR 2024
`-`	4D Gaussian Splatting for Real-Time Dynamic Scene Rendering	CVPR 2024
`CityDreamer`	CityDreamer: Compositional Generative Model of Unbounded 3D Cities	CVPR 2024
`Mip-Splatting`	Mip-Splatting: Alias-Free 3D Gaussian Splatting	CVPR 2024
`Text2NeRF`	Text2NeRF: Text-Driven 3D Scene Generation with Neural Radiance Fields	TVCG 2024
`Copilot4D`	Copilot4D: Learning Unsupervised World Models for Autonomous Driving via Discrete Diffusion	ICLR 2024
`OccWorld`	OccWorld: Learning a 3D Occupancy World Model for Autonomous Driving	ECCV 2024
`ReCamMaster`	ReCamMaster: Camera-Controlled Generative Rendering from A Single Video	arXiv 2025
`-`	Navigation world models	CVPR 2025
`PhysX-Anything`	PhysX-Anything: Simulation-Ready Physical 3D Assets from Single Image	arXiv 2025
`GAF`	GAF: Gaussian Action Field as a 4D Representation for Dynamic World Modeling in Robotic Manipulation	arXiv 2025
`ParticleFormer`	ParticleFormer: A 3D Point Cloud World Model for Multi-Object, Multi-Material Robotic Manipulation	arXiv 2025
`-`	3D and 4D World Modeling: A Survey	arXiv 2025
`GWM`	GWM: Towards Scalable Gaussian World Models for Robotic Manipulation	ICCV 2025
`MindCube`	MindCube: Spatial Mental Modeling from Limited Views	arXiv 2025
`GAIA-2`	GAIA-2: A Controllable Multi-View Generative World Model for Autonomous Driving	arXiv 2025
`-`	3D Reconstruction with Spatial Memory	3DV 2025
`-`	Continuous 3D Perception Model with Persistent State	arXiv 2025
`Fast3R`	Fast3R: Towards 3D Reconstruction of 1000+ Images in One Forward Pass	arXiv 2025
`WonderWorld`	WonderWorld: Interactive 3D Scene Generation from a Single Image	CVPR 2025
`-`	How to Enable LLM with 3D Capacity? A Survey of Spatial Reasoning in LLM	IJCAI 2025
`-`	Occupancy World Model for Robots	arXiv 2025
`TesserAct`	TesserAct: Learning 4D Embodied World Models	arXiv 2025
`Aether`	Aether: Geometric-Aware Unified World Modeling	ICCV 2025
`OccuBench`	OccuBench: Evaluating AI Agents on Real-World Professional Tasks via Language World Models	arXiv 2026
`PointWorld`	PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation	arXiv 2026
`CounterScene`	CounterScene: Counterfactual Causal Reasoning in Generative World Models for Safety-Critical Closed-Loop Evaluation	arXiv 2026
`MVISTA-4D`	MVISTA-4D: View-Consistent 4D World Model with Test-Time Action Inference for Robotic Manipulation	arXiv 2026
`X-scene`	X-scene: Large-scale driving scene generation with high fidelity and flexible controllability	NeurIPS 2026
`Code2World`	Code2World: A GUI World Model via Renderable Code Generation	arXiv 2026
`UniNav`	UniNav: A Unified World-Action Diffusion Model for Visual Navigation	arXiv 2026
`-`	SC $^{2}$ -WM: A Self-Correcting World Model with Closed-Loop Feedback for Vision-and-Language Navigation in Continuous Environments	arXiv 2026
`-`	Latent World Models with Monotone Planning Costs for Image-Goal Navigation	arXiv 2026
`Orbit`	Orbit-Planner: Towards Latent World Models for On-Orbit Obstacle Avoidance of Satellite Agents	arXiv 2026
`World`	World-Model-Grounded LLM Planning for AUV and ASV Navigation Near Offshore Wind Farms	arXiv 2026

Execution Proxy

What would the digital environment return after an operation? Web, GUI, code, shell, and API/tool execution predictors.

Model	Paper	Venue

`RT-2`	RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control	CoRL 2023
`-`	Code as Policies: Language Model Programs for Embodied Control	ICRA 2023
`Toolformer`	Toolformer: Language Models Can Teach Themselves to Use Tools	NeurIPS 2023
`-`	Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search	NeurIPS 2024
`AgentBench`	AgentBench: Evaluating LLMs as Agents	ICLR 2024
`-`	WorldCoder, a Model-Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment	NeurIPS 2024
`Math-Shepherd`	Math-Shepherd: Verify and Reinforce LLMs Step-by-Step without Human Annotations	ACL 2024
`-`	Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation	ICLR 2025
`CWM`	CWM: An Open-Weights LLM for Research on Code Generation with World Models	arXiv 2025
`-`	Web World Models	arXiv 2025
`WebSynthesis`	WebSynthesis: World-Model-Guided MCTS for Efficient WebUI-Trajectory Synthesis	arXiv 2025
`-`	Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents	TMLR 2025
`MegaSaM`	MegaSaM: Accurate, Fast, and Robust Structure and Motion from Casual Dynamic Videos	CVPR 2025
`ToolSandbox`	ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities	NAACL 2025
`ViMo`	ViMo: A Generative Visual GUI World Model for App Agents	arXiv 2025
`-`	Cosmos World Foundation Model Platform for Physical AI	arXiv 2025
`GTM`	GTM: Simulating the World of Tools for AI Agents	arXiv 2025
`-`	The Role of World Models in Shaping Autonomous Driving: A Comprehensive Survey	arXiv 2025
`GameFactory`	GameFactory: Creating New Games with Generative Interactive Videos	ICCV 2025
`ComAct`	ComAct: Reframing Professional Software Manipulation via COM-as-Action Paradigm	arXiv 2026
`MobileDreamer`	MobileDreamer: Generative Sketch World Model for GUI Agent	arXiv 2026
`Computer`	Computer-Using World Model	arXiv 2026
`MCP-Cosmos`	MCP-Cosmos: World Model-Augmented Agents for Complex Task Execution in MCP Environments	arXiv 2026
`-`	Generative Visual Code Mobile World Models	arXiv 2026
`IterCAD`	IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing	arXiv 2026
`-`	Debugging Code World Models	arXiv 2026
`NeuralOS`	NeuralOS: Towards Simulating Operating Systems via Neural Generative Models	ICLR 2026
`World-Model`	World-Model-Augmented Web Agents with Action Correction	arXiv 2026
`WebWorld`	WebWorld: A Large-Scale World Model for Web Agent Training	arXiv 2026
`AppDeltaWorld`	AppDeltaWorld: Transition-Grounded Delta Code World Model for Mobile GUI Agents	arXiv 2026
`WebRider`	WebRider: Persona-Conditioned Intent Controllers for Live-Web Assistance	arXiv 2026
`Twin`	Twin: Playing an Unknown Game with a Test-Time Digital Twin	arXiv 2026

Memory, Experience & Skill Proxy

What prior experience, constraints, or reusable skills does the agent need now? Store-retrieve memory systems and reusable skill/behavior libraries.

Model	Paper	Venue

`-`	Reasoning with Language Model is Planning with World Model	EMNLP 2023
`-`	Generative Agents: Interactive Simulacra of Human Behavior	UIST 2023
`-`	Cognitive Architectures for Language Agents	TMLR 2024
`Vismem`	Vismem: Latent vision memory unlocks potential of vision-language models	arXiv 2025
`MemHarness`	MemHarness: Memory Is Reconstructed, Not Replayed	arXiv 2026
`MemWM`	MemWM: Memory-Augmented Text-Based World Model	arXiv 2026
`-`	Activity Frames: Deterministic Screen-Activity Compilation for Agent Memory and Replay	arXiv 2026
`-`	Governed Persistent Memory: Source-Bound State Semantics and Fail-Closed Release for Long-Horizon Agents	arXiv 2026
`PBD`	PBD-AG: Persistent Baseline-Delta Active Graphs with Uncertainty-Aware Inspection for Long-Horizon Service Robots	arXiv 2026
`LUCID`	LUCID: Latent-Skill Unified Control via Imagined Dynamics for Long-Horizon Humanoid Loco-Manipulation	arXiv 2026
`-`	From Relevance to Execution Utility: Reward-Aware Dynamic Execution Gating for Skill-Based LLM Agents	arXiv 2026
`-`	Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation	arXiv 2026

Reward / Verification Proxy

Is the agent's behavior correct, safe, and feasible, and how should it improve? Reward models, verifiers, critics, preference models, and LLM-as-judge.

Model	Paper	Venue

`-`	Deep Reinforcement Learning from Human Preferences	NeurIPS 2017
`Fine`	Fine-Tuning Language Models from Human Preferences	arXiv 2019
`-`	Training Verifiers to Solve Math Word Problems	arXiv 2021
`-`	Direct Preference Optimization: Your Language Model is Secretly a Reward Model	NeurIPS 2023
`-`	Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena	NeurIPS 2023
`-`	A General Theoretical Paradigm to Understand Learning from Human Preferences	AISTATS 2024
`-`	A Survey on LLM-as-a-Judge	arXiv 2024
`ORPO`	ORPO: Monolithic Preference Optimization without Reference Model	EMNLP 2024
`-`	Let's Verify Step by Step	ICLR 2024
`-`	LLM Critics Help Catch LLM Bugs	arXiv 2024
`SimPO`	SimPO: Simple Preference Optimization with a Reference-Free Reward	NeurIPS 2024
`WorldSimBench`	WorldSimBench: Towards Video Generation Models as World Simulators	arXiv 2024
`WorldModelBench`	WorldModelBench: Judging Video Generation Models As World Models	arXiv 2025
`Computer`	Computer-Use Agents as Judges for Generative User Interface	arXiv 2025
`Critic-V`	Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning	CVPR 2025
`WCM`	WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning	arXiv 2026
`CheckVLA`	CheckVLA: Execution-Time Verification with Action-Conditioned World Model for Long-Horizon Mobile Manipulation	arXiv 2026
`ContactGuard`	ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models	arXiv 2026
`-`	Failure Detection for Surgical Robot Imitation Policies via Flow-Matching World Modeling	arXiv 2026
`-`	Calibrated Predictive Safety for Heterogeneous Robots: An Action-Conditioned JEPA Framework with Model-Based Safety Shields	arXiv 2026
`Ontology`	Ontology-Grounded World Models for Failure Diagnosis and Closed-Loop Repair in Physical AI Systems	arXiv 2026
`StructReward`	StructReward: Efficient Structured Process Rewards for Self-Correcting Multimodal Reasoning	arXiv 2026
`-`	Why Study Emergent Behavior When You Can Regulate It? Aligning Multi-Agent Systems with Reward Prediction	arXiv 2026
`-`	Refining Over Resampling: Test-Time Self-Correction for LLM Reasoning	arXiv 2026
`ARC`	ARC: Fair Relative Advantage Comparison in Open-Ended Real-World Interaction	arXiv 2026

Empowerment Levels (L1 / L2 / L3)

How proxies empower agents: L1 inference-time guidance (advise a frozen agent), L2 training-time optimization (generate experience/signal to train the agent), and L3 Agent-Proxy co-evolution (both improve each other in a loop).

Model	Paper	Venue

`-`	Scaling Agent Learning via Experience Synthesis	arXiv 2025
`RAGEN`	RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning	arXiv 2025
`VAGEN`	VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents	NeurIPS 2025
`WebEvolver`	WebEvolver: Enhancing Web Agent Self-Improvement with Coevolving World Model	EMNLP 2025
`-`	Aligning Agentic World Models via Knowledgeable Experience Learning	arXiv 2026
`-`	Agent World Model: Infinity Synthetic Environments for Agentic Reinforcement Learning	ICML 2026
`SearchGym`	SearchGym: Bootstrapping Real-World Search Agents via Cost-Effective and High-Fidelity Environment Simulation	arXiv 2026
`hint$^2$`	hint $^2$ : Hierarchical World Models for Inference-Time Temporal Logic Guidance	arXiv 2026

Evaluation, Fidelity & Safety

Benchmarks and metrics for world models, fidelity and calibration limits, and the safety/attack surface of proxy-driven learning.

Model	Paper	Venue

`G-Eval`	G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment	EMNLP 2023
`Length`	Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators	COLM 2024
`Prometheus`	Prometheus: Inducing Fine-Grained Evaluation Capability in Language Models	ICLR 2024
`BEHAVIOR-1K`	BEHAVIOR-1K: A Human-Centered, Embodied AI Benchmark with 1,000 Everyday Activities and Realistic Simulation	arXiv 2024
`-`	Evaluating the World Model Implicit in a Generative Model	NeurIPS 2024
`-`	World Models: The Safety Perspective	arXiv 2024
`-`	Understanding World or Predicting Future? A Comprehensive Survey of World Models	CSUR 2025
`-`	How Far is Video Generation from World Model: A Physical Law Perspective	ICML 2025
`-`	When World Models Dream Wrong: Physical-Conditioned Adversarial Attacks against World Models	arXiv 2026
`-`	Safety in Embodied AI: A Survey of Risks, Attacks, and Defenses	arXiv 2026
`JailWAM`	JailWAM: Jailbreaking World Action Models in Robot Control	arXiv 2026
`-`	Current Agents Fail to Leverage World Model as Tool for Foresight	arXiv 2026
`CtrlAttack`	CtrlAttack: A Unified Attack on World-Model Control in Diffusion Models	arXiv 2026
`SafeDream`	SafeDream: Safety World Model for Proactive Early Jailbreak Detection	arXiv 2026
`KineBench`	KineBench: Benchmarking Embodied World Models via IDM-Free Kinematic Grounding	arXiv 2026
`Apple`	Apple- $π$ : Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence	arXiv 2026
`WorldExam`	WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity	arXiv 2026
`GAUGE`	GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models	arXiv 2026
`H2R`	H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models	arXiv 2026
`WorldSimProbe`	WorldSimProbe: Diagnosing Simulator Faithfulness in Action-Conditioned World Models for Embodied Manipulation	arXiv 2026
`VIScore`	VIScore: Diagnosing Planning-Relevant Quality in Latent World Models	arXiv 2026
`HarnessEval`	HarnessEval-W: Agentifying the Evaluation of Visual Worlds	arXiv 2026
`CG`	CG-World: A Large-Scale World-State Dataset and Protocol for World Models	arXiv 2026
`-`	What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations	arXiv 2026
`-`	Ensuring Safe Physical AI in Urban Mobility via Hazard-Informed Synthesized Envelopes	arXiv 2026

Related surveys and the paper's core position statements.

Model	Paper	Venue

`-`	Is Sora a World Simulator? A Comprehensive Survey on General World Models and Beyond	arXiv 2024
`-`	A Survey of World Models for Autonomous Driving	arXiv 2025
`-`	A Comprehensive Survey on World Models for Embodied AI	arXiv 2025
`-`	Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond	arXiv 2026
`-`	Reinforcement Learning: From Algorithms To Foundation Models	arXiv 2026
`Planning`	Planning-Oriented End-to-End Autonomous Driving: Architectures, Evaluation, and Emerging Paradigms	arXiv 2026
`-`	Toward the Cognitive--Physical Limits of Embodied Intelligence through a World-Model-Centric Autonomous Racing Agent	arXiv 2026

Additional References

Further supporting works cited in the survey.

Model	Paper	Venue

`-`	Proximal Policy Optimization Algorithms	arXiv 2017
`Habitat`	Habitat: A Platform for Embodied AI Research	ICCV 2019
`-`	Model-Based Reinforcement Learning for Atari	ICLR 2020
`-`	Learning to Summarize from Human Feedback	NeurIPS 2020
`-`	Transporter Networks: Rearranging the Visual World for Robotic Manipulation	CoRL 2020
`CLIPort`	CLIPort: What and Where Pathways for Robotic Manipulation	CoRL 2021
`-`	Do As I Can, Not As I Say: Grounding Language in Robotic Affordances	CoRL 2022
`-`	Constitutional AI: Harmlessness from AI Feedback	arXiv 2022
`-`	Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback	arXiv 2022
`TensoRF`	TensoRF: Tensorial Radiance Fields	ECCV 2022
`-`	Inner Monologue: Embodied Reasoning through Planning with Language Models	CoRL 2022
`-`	Large Language Models are Zero-Shot Reasoners	NeurIPS 2022
`-`	Instant Neural Graphics Primitives with a Multiresolution Hash Encoding	TOG 2022
`-`	Training Language Models to Follow Instructions with Human Feedback	NeurIPS 2022
`Perceiver-Actor`	Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation	CoRL 2022
`-`	Solving Math Word Problems with Process- and Outcome-Based Feedback	arXiv 2022
`Chain-of`	Chain-of-Thought Prompting Elicits Reasoning in Large Language Models	NeurIPS 2022
`Plenoxels`	Plenoxels: Radiance Fields without Neural Networks	CVPR 2022
`Self`	Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture	CVPR 2023
`RT-1`	RT-1: Robotics Transformer for Real-World Control at Scale	RSS 2023
`Self-Refine`	Self-Refine: Iterative Refinement with Self-Feedback	NeurIPS 2023
`MemGPT`	MemGPT: Towards LLMs as Operating Systems	arXiv 2023
`Reflexion`	Reflexion: Language Agents with Verbal Reinforcement Learning	NeurIPS 2023
`ReAct`	ReAct: Synergizing Reasoning and Acting in Language Models	ICLR 2023
`-`	Tree of Thoughts: Deliberate Problem Solving with Large Language Models	NeurIPS 2023
`RRHF`	RRHF: Rank Responses to Align Language Models with Human Feedback without Tears	NeurIPS 2023
`KTO`	KTO: Model Alignment as Prospect Theoretic Optimization	ICML 2024
`SWE-bench`	SWE-bench: Can Language Models Resolve Real-World GitHub Issues?	ICLR 2024
`Oasis`	Oasis: A Universe in a Transformer	2024
`-`	Agent Planning with World Knowledge Model	NeurIPS 2024
`DeepSeekMath`	DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models	arXiv 2024
`VGGT`	VGGT: Visual Geometry Grounded Transformer	arXiv 2025
`MonST3R`	MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion	ICLR 2025
`-`	The latent space: Foundation, evolution, mechanism, ability, and outlook	arXiv 2026

Acknowledgements

This list is maintained alongside the survey. Thanks to all contributors and to the authors of the works listed here.

Citation

If you find this work useful, please consider citing:

@article{yang2026worldproxy,
  title   = {Quo Vadis, World Modeling?},
  author  = {Yu Yang and Xuemeng Yang and Licheng Wen and Lingdong Kong and Xiaobin Hu and Dongyue Lu and Wei Chow and Xiyan Huang and Yuxiang Feng and Yue Liao and Jianbiao Mei and Daocheng Fu and Rong Wu and Pinlong Cai and Ran Yi and Ying Tai and Jiangning Zhang and Botian Shi and Yong Liu and Shuicheng Yan},
  journal = {arXiv preprint arXiv:2608.02713},
  year    = {2026}
}