World models are one of the AI frontiers. Associated with 'generative 3D', games and robotics, they're systems that learn to predict abstract future states rather than generate plausible text.

A cluster of research threads, arriving from robotics, climate science, game design, and literary studies simultaneously, has converged on an unexpected hypothesis: narrative structure is a state-space problem, and the architecture being built to simulate physics is the same architecture needed to simulate story.

What World Models Actually Do

A large language model predicts the next token. A JEPA-style world model (the architecture championed by Yann LeCun ) predicts the next latent state: a compact, abstract representation of how the world has changed.

It never generates pixels or words during planning. It operates entirely in structured state space, then hands off to a separate renderer when output is needed.

Fei-Fei Li has formalised this into a three-tier taxonomy.

Renderers generate perceptually plausible outputs — video, images, spatial environments.

Simulators predict how states evolve over time under physical or causal laws.

Planners evaluate multiple possible futures and select trajectories that optimise for a goal.

Most AI content generation today lives at the Renderer tier.

The research frontier is Simulators and Planners, the space where narrative becomes relevant.

LeCun's $1 Billion Bet

In March 2026, Yann LeCun left Meta and closed a $1.03 billion seed round for Advanced Machine Intelligence Labs (AMI Labs), backed by NVIDIA, Samsung, Toyota Ventures, Bezos Expeditions, Jeff Bezos, and Eric Schmidt. Valued at $3.5 billion, it is Europe's largest seed round on record. AMI is building world models on the JEPA framework.

It is one of the most capitalised AI research programmes in existence, staking its entire thesis on the claim that latent-state prediction, not token generation, is the path to machine intelligence.

A concurrent paper, Causal-JEPA, explores the architecture further: it extends JEPA with object-level causal interventions, teaching world models not just that B follows A but that action A by agent X causes state change B.

Correlation versus causation at the event level is what separates a language model from a narrative simulator.

The Narrative Research Stack Is Already Being Built

Across four separate research programmes, the components of a story world model now have working prototypes.

Social dynamics as world state.

A joint CMU/NVIDIA paper introduced the Structured Social Simulation Analysis Protocol (S³AP), which represents social interactions as structured tuples of state, observation, agent actions, and mental states.

Tested on the FANToM theory-of-mind benchmark, it produced a 51% improvement in performance. The system predicts future social dynamics from story-extracted world states. This is the Simulator tier applied to human interaction.

Emotional arcs as generative constraints.

All Stories Are One Story implements Reagan et al.'s empirically derived six emotional arc types (the fundamental shapes that underlie most English-language fiction) as a directed acyclic graph of emotional states. Each node is automatically populated with characters, items, and plot-relevant attributes. Player evaluations confirmed that emotional arc integration significantly improves narrative coherence and engagement. The arc taxonomy, in other words, functions as a working prior for a generative world model.

Counterfactual literary simulation.

Matthew Wilkens at Cornell has made the explicit case — with early experimental evidence — that generative AI can function as a controlled simulation instrument for literary history.

The analogy he reaches for is climate science: "Climate scientists build computational models of the atmosphere to test the effects of greenhouse gas concentrations that have not yet occurred... When direct experimentation is impossible, simulation provides a principled alternative."

His paper calls for full counterfactual literary-historical simulations as explicit future work. A world model trained on literary history to answer questions like: what would the development of modernism have looked like without Joyce?

Earth system modelling as structural precedent.

The KIT "WOW" project in Germany is building an AI world model of the entire Earth climate system, funded at €6 million, by training on historical observational data and coupling specialised sub-models through shared latent representations.

The architecture (historical data, coupled sub-models, shared latent space, counterfactual simulation) maps directly onto what a narrative world model would require: sub-models for character psychology, plot tension, and world-state, coupled through shared representations of story state, trained on historical narrative corpora.

What the Architecture Looks Like

Synthesising Li's taxonomy and LeCun's framework, the components of a narrative world model have clear precedents for each layer.

A Story State in latent space encodes character beliefs, desires, and intentions; world facts; tension levels; and position on the emotional arc.

A Simulator — JEPA-style, trained on historical narrative data — predicts how that state evolves given a character action or plot event. A

Planner evaluates multiple possible story trajectories in imagination, selecting paths that optimise for narrative properties: coherence, arc completion, satisfying resolution.

A Renderer — potentially a conventional language model — generates surface text from the planned state.

Critically, this model never generates prose during planning. It operates in structured story-state space, the same way DreamerV3 and LeWorldModel operate in physical state space during control tasks. The text is a final output layer, not the reasoning medium.

Where the Gaps Are

No one has assembled these components into a complete story world model yet. The key open problems are substantial.

Whether JEPA-style architectures can be trained on sequential narrative data, rather than physical sensor data, remains theoretically plausible but practically undemonstrated at scale.

The right state representation for a story world has partial answers (S³AP tuples, emotional arc graphs) but no unified standard. Reagan et al.'s six-arc taxonomy was derived from English-language fiction; how far it generalises across non-Western narrative traditions is genuinely uncertain. And the interpretability problem is acute: most world model latent spaces are not human-readable, which limits their usefulness as a tool for working writers.

What to Watch

AMI Labs' first research publications will signal whether JEPA extends to non-physical domains. Any results on social or symbolic state spaces are directly applicable to narrative. Cornell's counterfactual literary simulation programme is the closest thing to a funded academic effort building the specific system described here.

The Causal-JEPA line of research is the architectural missing piece for story events. And World Labs, Fei-Fei Li's company, has moved past the Renderer tier; any announcement of non-spatial simulation capability would be a landmark.

The gap between what exists and what a narrative world model requires is narrowing. The pieces are in different labs, framed under different problem statements, funded for different applications. The convergence is visible. What does not yet exist is the assembly.

Sources include Fei-Fei Li's functional taxonomy, the Futurum Group analysis of AMI Labs' raise, Causal-JEPA, LeWorldModel, Social World Models — CMU/NVIDIA, All Stories Are One Story, Cornell's literary simulation paper, the KIT WOW project, and Latent.Space's adversarial reasoning analysis.