Language World Models (LWMs)

• Language World Models are language-based systems that represent and simulate environment dynamics, task transitions, and rewards using textual data.
• They employ varied methodologies including stateful simulation, implicit text prediction, and explicit symbolic decomposition to capture world states.
• LWMs enhance agent training and control by providing robust, evaluable simulation environments and supporting advanced interpretability studies.

Language World Models (LWMs) are world models implemented, conditioned, or expressed through language. In current research, the term denotes several closely related objects: a LLM configured as a stateful environment simulator for tool-using agents, a text-based transition model that predicts next observations or rewards from interaction history, an explicit symbolic or hierarchical world model specified in natural language, and a hypothesis about structured latent world representations inside LLMs themselves (Hu et al., 13 Apr 2026, Zuo et al., 23 Jun 2026, Feng et al., 2024). What unifies these usages is the treatment of environment dynamics, task structure, or latent world state as something that can be represented and manipulated in textual or language-mediated form.

1. Terminological scope and research usages

The literature does not use “Language World Model” in a single narrow sense. In some papers, the LWM is an external simulator that produces structured observations for an agent; in others, it is an implicit predictor of text-environment dynamics; in still others, it is an explicit symbolic scaffold or an internal representational hypothesis about LLMs.

Usage in the literature	Characteristic formulation	Representative papers
Stateful environment simulator	LLM “plays the role” of a domain tool-server	(Hu et al., 13 Apr 2026)
Implicit text-based transition model	Predicts next state, reward, or observation from history and action	(Li et al., 21 Dec 2025, Zuo et al., 23 Jun 2026)
Explicit symbolic or hierarchical world model	Preconditions/effects, H-MDP subgoals, or probabilistic programs	(Xie et al., 2024, Hill, 5 Sep 2025, Wong et al., 2023)
Latent internal world model	Structured world state decoded or probed from activations	(Feng et al., 2024, Tehenan et al., 3 Jun 2025, Robertson et al., 21 Jul 2025)

In "From Word to World: Can LLMs be Implicit Text-based World Models?" (Li et al., 21 Dec 2025), the LWM is a model $\mathcal{W}$ that predicts a text next state $S_n'$ and binary reward $R_n'$ from an initial state and an action history. In "Making LLMs into World Models with Precondition and Effect Knowledge" (Xie et al., 2024), the world model is decomposed into a precondition predictor and an effect predictor. In "Language-Driven Hierarchical Task Structures as Explicit World Models for Multi-Agent Learning" (Hill, 5 Sep 2025), the LWM is an explicit, hierarchical representation of task structure generated and maintained via natural-language reasoning. By contrast, "LLM world models are mental" (Robertson et al., 21 Jul 2025) uses the phrase to denote internal representations of physical systems that may resemble mental models.

This multiplicity is not merely terminological. It marks a genuine split between LWMs as deployable simulators for agents, LWMs as learned transition models for model-based RL, and LWMs as an interpretability or cognitive-science hypothesis about what LLMs encode internally.

2. Formalizations of language-mediated world dynamics

A canonical simulator-centric formalization appears in OccuBench, where an LWM is defined as

$(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$

with $s_t$ the latent environment state maintained implicitly in the context window, $a_t$ the agent’s tool call, $o_{t+1}$ the returned tool response, and $$c=(\text{system\_prompt}, \text{tool\_schema}, \text{initial\_state}, \text{state\_description})$$ the fixed configuration that defines the simulated domain. The same work also writes the simulator as a joint generative distribution,

$$P_{\theta}(s_{0:T}, o_{1:T}\mid a_{1:T}; c)=\prod_{t=0}^{T-1} P_{\theta}(s_{t+1}, o_{t+1}\mid s_t, a_t; c),$$

and emphasizes deterministic or low-temperature sampling so that each call yields a single coherent successor state and observation (Hu et al., 13 Apr 2026).

A text-environment formalization appears in the implicit text-based setting:

$$\mathcal{W}:\{S_0,(A_i,S_i')_{i=1}^{n-1},A_n\}\longrightarrow(S_n',R_n').$$

Here the model approximates the true environment dynamics in text space and is trained by minimizing cross-entropy on real trajectories, with evaluation by exact match or word-level $S_n'$0 on held-out transitions (Li et al., 21 Dec 2025). Qwen-AgentWorld states the same family of problems as next-observation prediction, defining an LWM whose sole objective is to predict the next environment observation from the system prompt, prior observations, and prior actions, i.e. $S_n'$1 (Zuo et al., 23 Jun 2026).

A more interpretable symbolic formulation decomposes transition modeling into action applicability and action effects. Xie et al. define a precondition predictor $S_n'$2 and an effect predictor $S_n'$3, then compose them into

$S_n'$4

with validity determined by whether the generated preconditions are semantically covered by the current state (Xie et al., 2024). This makes the learned world model inspectable at the level of atomic propositions rather than only token sequences.

Other formalisms extend beyond transition prediction. Language-Guided World Models condition the one-step model on a text manual $S_n'$5 and approximate

$S_n'$6

where $S_n'$7 is the trajectory history (Zhang et al., 2024). "From Word Models to World Models" instead treats meaning construction as a distribution over probabilistic programs $S_n'$8, with each program defining a generative model over latent world states $S_n'$9, so that

$R_n'$0

In that view, language is a route into a probabilistic language of thought rather than merely an interface to a simulator (Wong et al., 2023).

3. Architectures, synthesis pipelines, and training regimes

OccuBench operationalizes an LWM with four architectural components: a system prompt encoding domain logic and invariants, a JSON-style tool schema, an initial structured JSON state, and a human-readable state description that guides causal and constraint consistency across turns (Hu et al., 13 Apr 2026). To build the benchmark, it uses a multi-agent synthesis pipeline with a world-model synthesizer, a task-creator, and a verifier. For each scenario, the pipeline constructs a professional reference document; generates the environment configuration; proposes a task and canonical solution plan; derives an automated verification rubric; runs solvability checks and difficulty calibration; and applies a diversity filter based on 16 disjoint subtopics per scenario. The result is 382 evaluation instances across 100 scenarios.

Several papers treat LWMs as models that can be trained like next-token predictors and then post-trained for better transition fidelity. RLVR-World begins from maximum-likelihood autoregressive world models and then replaces purely likelihood-based optimization with reinforcement learning with verifiable rewards, using decoded transition metrics such as accuracy or $R_n'$1 as rewards. Its post-training pipeline is explicitly SFT $R_n'$2 RLVR, with Group Relative Policy Optimization, sampled groups of $R_n'$3 responses, and reward functions computed from decoded next-state predictions rather than from a learned reward model (Wu et al., 20 May 2025).

Qwen-AgentWorld scales this logic to foundation-model training for agentic environment simulation. Its three-stage pipeline is CPT $R_n'$4 SFT $R_n'$5 RL: continual pre-training injects general-purpose world modeling capabilities from state transition dynamics and augmented professional corpora; supervised fine-tuning activates next-state-prediction reasoning with explicit chain-of-thought; and reinforcement learning sharpens simulation fidelity through hybrid rubric-and-rule rewards. The training data comprise more than 10M environment interaction trajectories across 7 domains, and the RL reward combines a 5-dimensional rubric—Format, Factuality, Consistency, Realism, Quality—with deterministic rule-based verifiers (Zuo et al., 23 Jun 2026).

Not all LWM construction relies on sequence-only formulations. Xie et al. bootstrap synthetic training data for cooking plans by a global–local prompting pipeline: initial plan generation, local pruning and augmentation, precondition/effect inference, local re-generation, and global filtering. They then fine-tune two separate FLAN-T5-large models, one for preconditions and one for effects (Xie et al., 2024). In LIMT, the world model is conditioned on a frozen Sentence-BERT task embedding and trained jointly with an actor–critic in latent imagination, giving a language-informed multi-task visual world model rather than a purely text-based one (Aljalbout et al., 2024).

4. Evaluation protocols and benchmark design

OccuBench evaluates agent performance under both normal and faulty environments. It defines Completion Rate (CR) as the fraction of 382 tasks for which the full tool-call trajectory passes the verifier’s rubric, and Robustness Score

$R_n'$6

the worst relative drop under explicit, implicit, or mixed faults. Fault injection is parameterized by fault count $R_n'$7 and fault duration $R_n'$8, both with default value 2. The fault taxonomy distinguishes overt errors such as “HTTP 500” and “TimeoutError” from implicit degradations such as truncated lists, missing JSON fields, stale values, or null entries (Hu et al., 13 Apr 2026).

A broader evaluation framework for text-based LWMs separates fidelity and consistency, scalability and robustness, and agent utility. One-step fidelity is measured by exact match or word-level $R_n'$9 on next-state and reward prediction; long-horizon rollout consistency is measured by real success, world-model success, W2R replay success, and a Consistency Ratio

$(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$0

This framework is applied across ALFWorld, SciWorld, TextWorld, WebShop, and StableToolBench, and is meant to determine not only whether a world model predicts transitions accurately, but also whether it remains useful over rollouts and across environment complexity regimes (Li et al., 21 Dec 2025).

AgentWorldBench evaluates foundation LWMs as environment simulators for general agents. It contains 2,170 turn-level samples across 7 domains—MCP, Search, Terminal, SWE, Android, Web, and OS—constructed from real interactions of 5 frontier models on 9 established benchmarks. Evaluation is reference-grounded and uses a five-dimensional rubric scored by GPT-5.2 at turn level, with tokens split into deterministic, pre-existing, and runtime-metadata content types (Zuo et al., 23 Jun 2026).

These protocols reflect a significant methodological shift. Earlier world-model evaluation often centered on intrinsic next-state accuracy; the newer LWM literature increasingly insists on solvability, rollout consistency, robustness under controlled corruption, and downstream utility for agent learning or safety interventions.

5. Empirical findings, capability profiles, and failure modes

Across professional-task simulation, OccuBench reports four central findings: no single model dominates all industries; implicit faults are harder than both explicit errors and mixed faults; larger models, newer generations, and higher reasoning effort consistently improve performance; and strong agents are not necessarily strong environment simulators. A specific result is that GPT-5.2 improves by 27.5 points from minimal to maximum reasoning effort. On simulator fidelity, GPT-5.2 as a simulator produced an average CR of only 29.3%, versus 67.9% for Gemini-Flash, while Qwen 3.5 Plus reached 85.7% pairwise ranking agreement with Gemini-Flash and matched the top-3 agents exactly (Hu et al., 13 Apr 2026).

Internal-world-model evidence is more mixed. In pulley-system reasoning, state-of-the-art models performed significantly above the 20% chance baseline in mechanical-advantage estimation, at approximately 23–26% correct, and their estimates correlated with true mechanical advantage. However, follow-up analyses showed that number of pulleys was the strongest predictor of the model’s estimate, indicating a pulley-counting heuristic. The same models distinguished functional from randomly jumbled systems with mean $(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$1, but when non-functional systems were reconnected so that superficial connectedness was preserved while force transfer was broken, performance collapsed to chance, with mean $(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$2 and GPT-4o at $(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$3 (Robertson et al., 21 Jul 2025).

Interpretability studies provide stronger evidence for structured latent states than for reliable causal reasoning over those states. Propositional probes decode a latent world state from activations by binding token-level domain values into propositions such as $(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$4. In a closed-world setting, domain probes achieved more than 95% exact accuracy on synthetic data and approximately 90% or higher on paraphrase and translation; propositional probes achieved Jaccard approximately 0.9 across synthetic, paraphrase, and translated sets, and remained faithful under prompt injection, backdoor attacks, and gender-bias settings even when model outputs became unfaithful (Feng et al., 2024).

Spatial probing studies likewise report structured internal geometry. "Linear Spatial World Models Emerge in LLMs" finds near-perfect linear relation classification, inverse-relation alignment with cosine approximately 0.9969 for left–right and approximately 0.9950 for in front–behind in PCA space, 2D compositionality with mean cosine similarity approximately 0.9931, and overall causal steering success of 74.3% at layer 24, though success varied sharply by relation, from 100% for above, below, and left to 5% for behind (Tehenan et al., 3 Jun 2025). A plausible implication is that latent spatial structure can be linear and causally usable without yielding uniformly robust downstream reasoning.

6. Applications, significance, and open problems

One major application of LWMs is environment provision in domains where no real APIs or public testbeds exist. OccuBench explicitly frames this as turning the “untestable majority” of professional domains into a configurable prompt-engineering problem rather than an engineering problem of building real APIs (Hu et al., 13 Apr 2026). A second application is agent improvement through simulation: in text-based environments, action verification before irreversible actions yields +5–15 points for medium-capacity agents with $(s_{t+1}, o_{t+1}) = f_{\theta}(s_t, a_t; c),$5–10 simulations; synthetic trajectory generation gives Syn1K approximately equal to Real1K in SciWorld, with similar trends in WebShop; and warm-start RL on ALFWorld and SciWorld yields +10–15% higher success at early epochs and +5–10% final gains (Li et al., 21 Dec 2025).

A closely related line uses LWMs for AI control and human-in-the-loop revision. In Messenger, a Language-Guided World Model can imagine a full trajectory before acting, present that plan to a human, and update either the policy or the world model from action corrections or language feedback. The EMMA mechanism substantially outperforms both an observational baseline and a standard Transformer under compositional generalization splits, and in downstream return it approaches OracleParse across NewCombo, NewAttr, and NewAll settings (Zhang et al., 2024).

LWMs are also used as knowledge-transfer devices. In embodied simulation, fine-tuning LLMs with VirtualHome experience improves 18 downstream tasks by 64.28% on average, while EWC-LoRA preserves pre-training perplexity within 2–3% of the base model on held-out Pile data (Xiang et al., 2023). In multi-task robot learning, LIMT reaches a 52.1% multi-task success rate, versus 22.5% for MT-SAC:raw and 23.0% for MT-SAC:token, indicating that semantically meaningful task embeddings can improve parameter sharing in model-based multi-task learning (Aljalbout et al., 2024).

The open problems are now well delineated. Reward design remains crucial in RL-based world-model post-training; out-of-distribution generalization and counterfactual reasoning remain open challenges; high-entropy or partially observable environments produce rollout drift; and simulator reliability must be audited before benchmark conclusions are trusted (Wu et al., 20 May 2025, Li et al., 21 Dec 2025). Other proposed directions include retrieval-augmented world models, multi-modal or structured latent states, explicit physics engines or structured world-model modules, and architectural separation between world understanding and linguistic realization, as in the “mouth is not the brain” principle of Boltzmann-GPT (Niimi, 23 Jan 2026, Robertson et al., 21 Jul 2025).

Taken together, the LWM literature no longer supports a single thesis about what LLMs “really are.” It instead defines a technical program spanning simulator construction, transition modeling, symbolic decomposition, latent-state interpretability, and control. The central empirical lesson is not that language alone suffices for robust world modeling, but that language provides a versatile substrate on which diverse forms of world modeling can be built, evaluated, and, in many domains, made operational.