Agent-World: World Models & Simulation

• Agent-World is a research paradigm where agents utilize explicit internal and external world representations—such as cognitive maps, executable simulations, and semantic models—to mediate observation, action, and learning.
• This approach employs diverse methodologies, including DOM transduction, tool-using reinforcement learning, and language world models, to construct and update dynamic representations.
• Agent-World techniques enhance planning, foresight, and multi-agent collaboration by enabling agents to simulate, verify, and scale interactions with higher fidelity and adaptability.

Agent-World denotes a line of research in which autonomous behavior depends on an explicit world representation or world substrate that mediates between observation, action, and learning. In recent work, that world may be an internal cognitive map built from DOM trees and semantic service descriptions, a text or language world model that predicts next observations, a fully synthetic executable environment for tool-using reinforcement learning, a shared semantic world for multi-agent systems, or a multi-view generative model that keeps multiple agents synchronized to one underlying scene (Gidey et al., 28 Oct 2025, Li et al., 8 Jun 2026, Wang et al., 10 Feb 2026, Mantsivoda et al., 1 Apr 2026, Zhu et al., 3 Mar 2026). This suggests that Agent-World is not a single architecture but a family of approaches organized around how agents construct, simulate, share, and update the worlds in which they operate.

1. Conceptual scope and definitions

The literature uses the term in several precise senses. In one strand, the “world” is an internal actionable representation, such as a cognitive map or world model that fuses perceptual streams into a task-usable structure. In another, the “world” is an executable environment whose state, actions, observations, and rewards are explicitly specified. In a third, the “world” is a shared semantic substrate that multiple agents inhabit and modify. In a fourth, the “world” is a synchronized generative process that must remain consistent across multiple agents and viewpoints.

Sense of Agent-World	Canonical object	Representative sources
Internal actionable representation	cognitive map / world model, Page Affordance Model, affordance catalog	(Gidey et al., 28 Oct 2025)
Executable training world	POMDP with database state and tool-mediated actions	(Wang et al., 10 Feb 2026, Dong et al., 20 Apr 2026)
Shared explicit world	world-centered multi-agent system $W=(E,R,S,A,T,C)$	(Mantsivoda et al., 1 Apr 2026)
Shared generative world	synchronized multi-agent videos or multi-view latent world state	(Zhu et al., 3 Mar 2026, Wu et al., 20 Apr 2026, Liu et al., 27 May 2026)
Language world model	next-observation predictor or text world model	(Zuo et al., 23 Jun 2026, Li et al., 8 Jun 2026)

A recurring misconception is that a world model must always be a latent dynamics module in the model-based RL sense. The current literature is broader. It includes explicit semantic ontologies, hierarchical task graphs, executable simulators, symbolic domains, and text-rendered transition models, in addition to learned latent predictors (Mantsivoda et al., 1 Apr 2026, Hill, 5 Sep 2025, Hu et al., 26 Dec 2025).

2. World construction from structured digital environments

A central Agent-World formulation treats web and service interaction as a world-model construction problem rather than direct action on raw structure. In this formulation, raw HTML and API descriptions are not yet agent-usable worlds. Raw DOM is highly verbose, contains scripts, styles, trackers, and boilerplate, and cited pruning studies suggest that 80–90% may be non-semantic; hardcoded API integrations are brittle because they couple the agent to a particular service version, break under change, and prevent runtime discovery (Gidey et al., 28 Oct 2025).

The proposed remedy is a pattern language with two complementary architectural patterns. The DOM Transduction Pattern inserts a DOM Transformer into perception. It ingests a structured DOM tree and outputs a Page Affordance Model (PAM) through cleaning, pruning, compact representation, and optionally an LLM-as-Transformer stage. The explicit requirement is that PAM preserve the essential structure of task-relevant interactive elements while remaining decoupled from the reasoning core. The Hypermedia Affordances Recognition Pattern addresses unknown services and devices by parsing standardized semantic descriptions, most concretely W3C Web of Things Thing Description (TD), extracting Properties, Actions, Events, and protocol bindings, and updating the cognitive map with an affordance catalog discovered at runtime (Gidey et al., 28 Oct 2025).

This view aligns closely with the formalization of text world models (TWMs), where a transition model predicts the next textual state from the current state and action,

$$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$

and the successor state may be natural language, JSON, key-value records, accessibility trees, PDDL-style predicates, or executable code (Li et al., 8 Jun 2026). The survey literature further organizes construction methods into LLM-as-world-model and code-as-world-model paradigms, separating implicit dynamics absorbed into weights from executable dynamics externalized into programs (Li et al., 8 Jun 2026).

The significance of this strand is representational rather than merely compressive. The objective is not just to shorten input, but to convert external structure into affordance-bearing internal state. This suggests that, in digital environments, much of the Agent-World problem is a problem of choosing the right intermediate world representation rather than a problem of direct end-to-end action selection.

3. Synthetic, executable, and self-evolving environments

Another major use of Agent-World concerns the construction of executable training worlds. Agent World Model (AWM) defines each synthetic environment as a POMDP,

$$E_i = (\mathcal{S}_{E_i}, \mathcal{A}_{E_i}, \mathcal{O}_{E_i}, T_{E_i}, \{R_\tau\}_{\tau \in \mathcal{T}_{E_i}}),$$

with the design principle database = state, MCP tool interface = action/observation layer, and verification code = reward construction (Wang et al., 10 Feb 2026). The pipeline scales to 1,000 environments and 10,000 tasks, backed by 35,062 tools total, with an average of 35.1 exposed tools per environment, 18.5 database tables, 129.3 sample data records, 8.5 agent steps per task, and 7.1 unique tools used per task; 13.7% of tasks exceeded the 20-step budget during complexity analysis (Wang et al., 10 Feb 2026). The stated motivation is that code-driven, database-backed environments provide more reliable transitions and cheaper repeated interaction than LLM-simulated environments.

A complementary line scales realistic rather than purely synthetic environments. Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence introduces two components: Agentic Environment-Task Discovery and Continuous Self-Evolving Agent Training (Dong et al., 20 Apr 2026). The environment ecosystem contains 1,978 retained environments and 19,822 distinct tools from over 2K themes, and training is evaluated on 23 benchmarks. The paper reports that increasing the number of training environments from 0 to approximately 2000 raises the average across four representative domains from 18.4% to 38.5%, and that self-evolution rounds yield monotonic gains with diminishing returns (Dong et al., 20 Apr 2026).

A third executable strand focuses not on environment supply for RL, but on generating world models themselves. Agent2World maps a text description $x$ to an executable symbolic world model

$$F(x)=WM,\qquad WM=(P_{\text{env}},A_{\text{env}},T_{\text{env}}),$$

with target formats including PDDL and Python (Hu et al., 26 Dec 2025). Its three-stage system comprises Deep Researcher, Model Developer, and a Testing Team that combines adaptive unit testing and simulation-based validation. Only repair trajectories whose final artifacts pass verification are kept for supervised fine-tuning, and the paper reports an average relative gain of 30.95% over the same model before training (Hu et al., 26 Dec 2025).

These systems treat the world not as a fixed benchmark but as an engineered object: synthesizable, verifiable, and revisable. A plausible implication is that environment quality becomes a first-class scaling axis for general agents, alongside model size and policy optimization.

4. Planning, foresight, and language world models

In planning-oriented work, Agent-World refers to predictive machinery that provides foresight before an action is executed. Agent Planning with World Knowledge Model formulates interactive planning as a POMDP and distinguishes prior task knowledge from dynamic state knowledge (Qiao et al., 2024). The World Knowledge Model (WKM) is a parametric knowledge model trained to self-synthesize knowledge from expert and sampled trajectories, then provide global prior knowledge for early planning and dynamic state knowledge for local planning. The paper reports that WKM reduces blind trial-and-error and hallucinated actions; on ALFWorld, invalid-action rates fall to 32.86% seen and 29.85% unseen, compared with higher rates for several baselines (Qiao et al., 2024).

WorldEvolver studies a more explicit foresight setting in which an LLM agent operates in a partially observed interaction process $(\mathcal{S},\mathcal{A},\mathcal{O},\mathcal{T})$, sees

$s_t=(o_1,a_1,\ldots,o_{t-1},a_{t-1},o_t),$

samples $a_t\sim\pi_\theta(\cdot\mid s_t)$, and queries a frozen world model for predicted future observations (Zhang et al., 29 Jun 2026). Its central claim is that foresight helps only when reliable. The framework therefore maintains a memory state

$M_t=(M_E^t,M_S^t),$

with Episodic Memory over realized transitions, Semantic Memory over mismatch-derived heuristic rules, and Selective Foresight that abstains when confidence is too low. In the reported planning results, Gemma-4-26B-A4B with ReAct improves from 23.88 to 26.12 on ALFWorld and from 44.44 to 52.22 on ScienceWorld when WorldEvolver with selective foresight is used (Zhang et al., 29 Jun 2026).

At foundation-model scale, Qwen-AgentWorld trains native language world models to predict next observations directly from long interaction histories,

$\hat{o}_{t+1}=f_\theta(c,o_{\le t},a_{\le t}),$

across 7 domains: MCP, Search, Terminal, SWE, Android, Web, and OS (Zuo et al., 23 Jun 2026). Training uses a three-stage recipe summarized as “CPT injects, SFT activates, RL sharpens”, built on more than 10M environment interaction trajectories and evaluated on AgentWorldBench, which contains 2,170 evaluation samples from 9 established benchmarks (Zuo et al., 23 Jun 2026). On the main five-dimensional rubric average, Qwen-AgentWorld-397B-A17B reaches 58.71 overall and Qwen-AgentWorld-35B-A3B reaches 56.39, with the larger model outperforming existing frontier models overall and showing particular strength on Terminal and SWE (Zuo et al., 23 Jun 2026).

The broader survey literature frames these developments as a response to the agent-world gap: many LLM agents remain reactive, whereas world models provide planning, verification, training-time experience synthesis, and principled evaluation (Li et al., 8 Jun 2026). The literature also stresses that high prediction accuracy does not necessarily imply high task utility, because a next-state predictor may miss the decision-critical information needed by the downstream agent (Li et al., 8 Jun 2026).

5. Explicit shared worlds, semantic models, and theoretical necessity

A more explicit and institution-oriented formulation appears in world-centered multi-agent systems (WMAS). Here the primary representational substrate is not each agent’s private belief state, but a shared world

$$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$0

where $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$1 is a set of entities, $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$2 a set of relations, $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$3 a state space over $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$4, $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$5 admissible actions, $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$6 a transition function or relation, and $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$7 a set of constraints or norms (Mantsivoda et al., 1 Apr 2026). WMAS is motivated by structured domains such as enterprises, finance, healthcare organizations, and institutional workflows, where explicit state control, semantic consistency, explainability, long-term stability, and verifiable behavior are required. Its semantic model has a ground semantic layer implemented as object ontologies and a causal knowledge layer updated by Semantic Machine Learning (SML); Ontobox is presented as a concrete realization (Mantsivoda et al., 1 Apr 2026).

In long-horizon multi-agent learning, explicit world structure is instead cast as a hierarchy of tasks and subgoals. Language-Driven Hierarchical Task Structures as Explicit World Models for Multi-Agent Learning argues that failure in robotic soccer arises because environments are too structurally flat, not because they are too low-fidelity (Hill, 5 Sep 2025). The proposed world model is a task graph, often a directed acyclic graph (DAG), in which tasks, sub-tasks, dependencies, and intrinsic rewards are first-class objects. The paper introduces benchmark-style metrics including Compositional Generalization Score (CGS), Curriculum Efficiency Gain (CEG) with

$$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$8

and Scaffolding Brittleness Index (SBI) (Hill, 5 Sep 2025). In this view, the environment becomes an active teacher.

Theoretical work strengthens the claim that world-model-like structure is not optional for competent agency. General Agents Contain World Models, even under Partial Observability and Stochasticity proves that sufficiently general, almost optimal agents encode enough information about their environments to permit approximate reconstruction of transition probabilities by black-box querying, even for stochastic agents in partially observable environments (Cifuentes, 3 Feb 2026). The result is behavioral rather than representational: it does not require an explicit symbolic model, but it does imply that competent generalized behavior contains recoverable world information. This suggests a bridge between explicit-world and implicit-world accounts: some Agent-World systems externalize the world, while others necessarily internalize it.

6. Embodied, visual, and multi-agent shared-world simulation

In visual and embodied settings, Agent-World often denotes a requirement that multiple observations correspond to one coherent underlying scene. ShareVerse introduces multi-agent shared world modeling for video generation, using a CARLA-based dataset with two synchronized agents, each equipped with four synchronized cameras—front, rear, left, and right—yielding 8 video streams total per sample (Zhu et al., 3 Mar 2026). The framework spatially concatenates four views per agent and adds cross-agent attention blocks to a pretrained CogVideoX backbone so that spatial-temporal information transfers across agents. It supports 49-frame generation and reports PSNR 20.76, SSIM 0.6656, and LPIPS 0.2791 (Zhu et al., 3 Mar 2026).

MultiWorld generalizes this setting to scalable multi-agent multi-view world modeling through a Multi-Agent Condition Module (MACM) and a Global State Encoder (GSE) (Wu et al., 20 Apr 2026). MACM provides identity-aware, interaction-aware conditioning over multiple agent actions, while GSE uses a pretrained VGGT backbone to encode a shared 3D-aware latent state from multiple views. On the reported multi-player game dataset, MultiWorld achieves FVD 179, Action 89.8, and RPE 0.67, outperforming several baselines on overall fidelity and multi-view consistency; on multi-robot manipulation it attains FVD 96 and RPE 1.52 (Wu et al., 20 Apr 2026). Gamma-World extends multi-agent world modeling beyond two players through Simplex Rotary Agent Encoding, Sparse Hub Attention, and a teacher-student pipeline enabling rollout at 24 FPS, with the reported ability to generalize from two to four players without additional training (Liu et al., 27 May 2026).

For model-based MARL, the shared world need not be pixel-based. DIMA models global state transitions in cooperative Dec-POMDPs through sequential revelation of agent actions, aligning uncertainty reduction with a diffusion-inspired reverse process over the next state (Zhang et al., 27 May 2025). It reports state-of-the-art performance on MAMuJoCo and Bi-DexHands, for example 4881 ± 756 on Ant-2x4 and 259.9 ± 4.1 on BottleCap (Zhang et al., 27 May 2025). MATWM instead uses a transformer-based world model with decentralized imagination, a semi-centralized critic, and a teammate prediction module, and reports strong sample efficiency, including near-optimal performance in as few as 50K environment interactions (Deihim et al., 23 Jun 2025).

Embodied simulation platforms expand the notion further. AgentWorld is an interactive simulation platform for household mobile manipulation that combines automated scene construction, VR-based teleoperation, imitation-learning data collection, and sim-to-real transfer (Zhang et al., 11 Aug 2025). It supports multiple embodiments including Unitree G1, Unitree H1, Franka Emika Panda, DOBOT X-Trainer, and TRX-Hand5, uses RGB observations at 480 × 640, and defines hybrid mobile-manipulation state and action spaces over arm, end-effector, floating base, and a binary locomotion/manipulation mode selector (Zhang et al., 11 Aug 2025). SimWorlds addresses dynamic 4D scene creation in Blender through a planner-coder-reviewer workflow, a fixed ordered stage sequence, a layered scene protocol, and deterministic verification. Its 4DBuildBench contains 50 scenes—45 dynamic and 5 static—and SimWorlds reports MPR 0.87, SPR 0.89, and VLM score 0.82, outperforming VIGA on mechanism and structural correctness (Liu et al., 2 Jul 2026).

A different embodied multi-agent perspective appears in Multi-Agent Craftax (MAC), an efficient open-world environment for social learning. MAC runs 100 million training steps in under one hour on a single GPU, uses 4 agents in the reported experiments, and studies whether agents learn from experts, tool sharing, cooperation, or competition (Ye et al., 21 Aug 2025). The salient negative result is that social learning from experts did not improve performance meaningfully: the cultural transmission score is reported as $$M : \mathcal{S} \times \mathcal{A} \to \mathcal{T}_{\mathcal{S}}, \qquad \hat{s}_{t+1}=M(s_t,a_t),$$9 without auxiliary loss and $$E_i = (\mathcal{S}_{E_i}, \mathcal{A}_{E_i}, \mathcal{O}_{E_i}, T_{E_i}, \{R_\tau\}_{\tau \in \mathcal{T}_{E_i}}),$$0 with auxiliary loss (Ye et al., 21 Aug 2025). This result complicates simple claims that richer worlds automatically yield richer learning.

7. Evaluation, limitations, and open problems

Evaluation in Agent-World research is heterogeneous because the “world” itself varies. The survey on text world models separates evaluation of the world model from evaluation using the world model as an environment, and notes that metrics such as exact match, token F1, consistency ratio, and behavior-consistency variants may disagree with downstream planning value (Li et al., 8 Jun 2026). This is echoed in system-specific studies. WorldEvolver reports that noisy foresight can hurt planning more than no foresight at all (Zhang et al., 29 Jun 2026). SimWorlds explicitly separates engine-state audit from itemized VLM judging because vision models are unreliable at detecting whether motion was produced by the correct physical mechanism (Liu et al., 2 Jul 2026). Agent2World argues that parseability and static validation are insufficient because behavior-level failures appear only during interactive execution (Hu et al., 26 Dec 2025).

The limitations are similarly varied. In the DOM/service setting, aggressive pruning can remove critical elements, and dynamic service discovery depends on correctly published Thing Descriptions (Gidey et al., 28 Oct 2025). In language world modeling, GUI domains remain weaker under text-centric simulation, and factuality is the lowest-scoring rubric dimension even after RL in Qwen-AgentWorld (Zuo et al., 23 Jun 2026). In multi-agent video generation, ShareVerse is limited to two agents and remains simulation-based, while a real-time, long-term consistent shared world model is still described as an open direction (Zhu et al., 3 Mar 2026). MultiWorld notes that large-scale training remains unexplored due to computational constraints and that small or distant agents can remain ambiguous because of limited spatial resolution (Wu et al., 20 Apr 2026). In social-learning worlds, MAC shows that current social-learning mechanisms do not yet reliably exploit expert presence or proximity (Ye et al., 21 Aug 2025).

Several open problems recur across strands. One is world model–policy coupling: whether the world model should be shared with the agent, decoupled as a simulator, or used only as a verifier or critic (Li et al., 8 Jun 2026). Another is adaptation under drift: websites change, APIs evolve, user behavior shifts, and world models must update without catastrophic forgetting or uncontrolled simulator drift (Zhang et al., 29 Jun 2026, Zuo et al., 23 Jun 2026). A third is mechanism fidelity: many tasks require not merely plausible outputs but correct state transitions, causal structure, or physically valid construction (Hu et al., 26 Dec 2025, Liu et al., 2 Jul 2026). A fourth is world scale: several papers report positive scaling with environment diversity, self-evolution rounds, or numbers of agents and views, but also identify diminishing returns and growing systems complexity (Dong et al., 20 Apr 2026, Liu et al., 27 May 2026).

Taken together, these works suggest a common thesis. Agent capability is increasingly being treated not as policy competence alone, but as the joint ability to build, query, inhabit, and revise a world. Whether the world is internal, executable, shared, semantic, symbolic, or generative, the contemporary Agent-World literature treats world construction and world maintenance as core problems of autonomous intelligence rather than auxiliary implementation details.