Centrality Zoo: Misnomer in World Models

• Centrality Zoo is a misnomer since the literature actually documents language world models for simulating and controlling dynamic environments.
• Research emphasizes explicit precondition and effect models along with latent representations to enhance planning and evaluation in simulated contexts.
• Studies showcase diverse methodologies—from symbolic simulation to reinforcement learning pipelines—with consistent improvements in world model fidelity.

Centrality Zoo is not a term defined or discussed in the cited arXiv materials. The source set instead documents a research program on Language World Models (LWMs): systems in which LLMs simulate environment dynamics, infer preconditions and effects, maintain latent world states, or serve as configurable environment simulators for planning and evaluation (Xie et al., 2024). Within the evidentiary boundary of these sources, the only rigorous treatment possible is therefore a delineation of that mismatch and a summary of the attested topic actually covered by the literature.

1. Attestation and scope

No paper in the cited set introduces a construct, benchmark, taxonomy, or formalism named “Centrality Zoo.” The referenced works are instead about language-conditioned or language-implemented world modeling, including explicit precondition/effect models, autoregressive next-state predictors, latent-state probes, embodied finetuning, agentic environment simulators, and language-guided control (Hu et al., 2023).

This absence matters methodologically. A term can support encyclopedia treatment only when the source base supplies a stable denotation, recurring usage, or at least a definitional anchor. Here, none is present. The attested subject matter is “world models,” not graph-theoretic centrality, a “zoo” of centrality measures, or any analogous named object.

A plausible implication is that the requested title has been mismatched to a source bundle on LWMs. That inference does not alter the documentary fact: the cited papers substantiate a literature on world modeling with LLMs, not on “Centrality Zoo.”

2. Conceptual nucleus of the documented literature

The most explicit symbolic formulation in the source set defines a discrete state space $S$, a finite action set $A$, a precondition inference function $f_{\mathrm{pre}}: A \to 2^L$, and an effect inference function $f_{\mathrm{eff}}: A \to (Add(a), Del(a))$, with valid-action prediction given by $Valid(a,s)=\mathrm{true} \iff Pre(a)\subseteq s$ and a STRIPS-style transition $T(s,a)=(s\setminus Del(a))\cup Add(a)$ (Xie et al., 2024). In that formulation, both preconditions and effects are represented as natural-language sentences, and applicability is checked by a separate semantic-matching LLM.

A broader planning-oriented abstraction appears in the LAW framework, where a world model is written as a stochastic transition kernel $\mathcal{T}(s' \mid s,a)$, and an agent model adds beliefs $b(s)$, rewards $R:\mathcal{S}\times\mathcal{A}\to\mathbb{R}$, and a policy $\pi(a\mid b)$ to support deliberate reasoning and search (Hu et al., 2023). This connects LLMs, agent models, and world models as distinct but composable elements.

Other papers generalize the same idea to text-native simulation. Qwen-AgentWorld defines an LWM as a conditional LLM

$A$0

where the model directly generates the next environment observation as text (Zuo et al., 23 Jun 2026). OccuBench likewise formalizes an LWM as

$A$1

with $A$2 containing a system prompt, tool schema, initial state, and state description (Hu et al., 13 Apr 2026). Across these formulations, the common core is not centrality analysis but language-mediated state transition modeling.

3. Explicit-state and explicit-knowledge approaches

One major strand of the literature makes world knowledge explicit rather than leaving it latent in hidden states. “Making LLMs into World Models with Precondition and Effect Knowledge” trains two separate FLAN-T5-large models: one maps an action to $A$3, and the other maps an action to $A$4, using a synthetic corpus generated by a “global-local prompting” pipeline with GPT-4 (Xie et al., 2024). The resulting corpus contains $A$5 plans with $A$6 steps on average in the cooking domain, and the full method improves both precondition and effect inference relative to ablations without global or local pruning.

A related but more general control-oriented formulation appears in “Language-Guided World Models: A Model-Based Approach to AI Control,” where an LWM is a one-step probabilistic model $A$7 conditioned on a textual manual $A$8 (Zhang et al., 2024). Its EMMA mechanism separates entity identification from attribute extraction and is designed specifically for compositional generalization in the Messenger environment.

The probabilistic-programming line pushes explicitness further. “From Word Models to World Models: Translating from Natural Language to the Probabilistic Language of Thought” treats linguistic meaning as a mapping from utterances $A$9 to a distribution over probabilistic programs $f_{\mathrm{pre}}: A \to 2^L$0, with downstream beliefs obtained by marginalizing over programs: $f_{\mathrm{pre}}: A \to 2^L$1 (Wong et al., 2023). In that framework, world models are executable symbolic substrates rather than merely neural predictors.

These approaches are unified by an insistence that a usable world model should expose structured state, action, or causal content. None of them is presented as a taxonomy of centrality measures.

4. Latent-state and emergent-world-model approaches

A second strand argues that world models can exist implicitly inside model activations. “Monitoring Latent World States in LLMs with Propositional Probes” defines a latent world state as an internal representation encoding propositions such as $f_{\mathrm{pre}}: A \to 2^L$2, then introduces domain probes and a learned “binding subspace” to recover compositional propositions from activations (Feng et al., 2024). In prompt injection, backdoor, and gender-bias settings, the decoded propositions remain more faithful than the model’s output text, suggesting a divergence between internal state and external decoding.

“Linear Spatial World Models Emerge in LLMs” supplies a geometric variant of the same claim. It defines a spatial world model as $f_{\mathrm{pre}}: A \to 2^L$3, requires basis and composition properties for spatial relations, and reports that linear probes recover object positions while causal steering along relation vectors changes spatial outputs (Tehenan et al., 3 Jun 2025). The paper interprets this as evidence for a linear spatial embedding isomorphic to $f_{\mathrm{pre}}: A \to 2^L$4.

The philosophical counterpart appears in “From task structures to world models: What do LLMs know?”, which distinguishes “instrumental knowledge” from “worldly knowledge” and characterizes world models as homomorphic, structure-preserving abstractions of causal reality (Yildirim et al., 2023). On that view, what matters is not only task performance but whether the model recovers structured representations analogous to intuitive physics engines, cognitive maps, or embodied planners.

The mechanical-reasoning study “LLM world models are mental” sharpens the limit case. It finds above-chance performance on pulley tasks, significant correlations between estimates and true mechanical advantage, and sensitivity to gross functional structure, but near-chance behavior when subtle force-transmission connectivity must be represented (Robertson et al., 21 Jul 2025). The documented conclusion is not the absence of world models but their brittleness.

5. Agentic simulators, training pipelines, and benchmarks

Several papers move from representation to deployment as general-purpose simulators. RLVR-World casts autoregressive next-state prediction as a policy over output tokens, then optimizes decoded next-state metrics with reinforcement learning using verifiable rewards and Group Relative Policy Optimization (Wu et al., 20 May 2025). In text games, RLVR-World improves overall accuracy from $f_{\mathrm{pre}}: A \to 2^L$5 under SFT to $f_{\mathrm{pre}}: A \to 2^L$6 or $f_{\mathrm{pre}}: A \to 2^L$7, depending on reward design; in WebArena-derived state prediction it raises F1 from $f_{\mathrm{pre}}: A \to 2^L$8 to $f_{\mathrm{pre}}: A \to 2^L$9.

Qwen-AgentWorld extends this to large-scale agentic simulation. It introduces Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, trained through a three-stage pipeline in which “CPT injects world knowledge,” “SFT activates next-state reasoning,” and “RL sharpens fidelity,” using more than 10M environment interaction trajectories over 7 domains (Zuo et al., 23 Jun 2026). Evaluation is organized through AgentWorldBench, a 2,170-sample, 7-domain suite built from real-world interactions on 9 established benchmarks.

OccuBench uses LWMs as domain simulators for professional work. Its benchmark covers 100 scenarios across 10 industry categories and 65 specialized domains, with evaluation along both task completion and robustness under explicit, implicit, and mixed faults (Hu et al., 13 Apr 2026). The paper also emphasizes that simulator quality is decisive for benchmark reliability: cross-simulator agreement can remain high in one setting yet degrade sharply when a capable agent functions poorly as a simulator.

A complementary embodied route appears in “LLMs Meet World Models: Embodied Experiences Enhance LLMs,” where a LLM is finetuned on VirtualHome interaction transcripts acquired through goal-oriented planning and random exploration, with EWC and LoRA used to preserve general language competence (Xiang et al., 2023). The reported outcome is an average improvement of $f_{\mathrm{eff}}: A \to (Add(a), Del(a))$0 across 18 downstream tasks.

These papers collectively constitute a substantial “zoo” only in the informal sense of methodological diversity across explicit symbolic models, autoregressive simulators, latent-state probes, and embodied finetuning. That is an editor’s characterization of the attested LWM literature, not a source-backed name for the topic “Centrality Zoo.”

6. Boundary conditions on interpretation

The source set supports several recurring conclusions about LWMs. First, fidelity and planning utility depend strongly on data quality, behavioral coverage, and domain structure (Li et al., 21 Dec 2025). Second, explicit structure—whether preconditions and effects, task graphs, tool schemas, or state descriptions—often improves controllability and evaluation (Xie et al., 2024). Third, emergent internal world models can be detectable yet brittle, especially when subtle structural constraints must be tracked (Robertson et al., 21 Jul 2025).

None of these findings licenses a substantive encyclopedia treatment of “Centrality Zoo” as such. No source here provides a definition, genealogy, benchmark, empirical comparison, or controversy under that title. The documented literature instead concerns Language World Models, their symbolic and latent forms, their training objectives, and their use in planning, simulation, interpretability, and benchmark construction (Li et al., 21 Dec 2025).

A plausible implication is that “Centrality Zoo” belongs to a different bibliography than the one cited here. On the evidence available, the only encyclopedically accurate statement is that the present source base does not attest that topic and instead maps a diverse but coherent research area centered on language-based world modeling.