The last few years have seen remarkable progress in foundation models. We now have models that write essays, generate code, summarize research papers, and reason over long contexts. Yet when asked to imagine the future, these same models reveal an important limitation: they are trained to continue text, not to generate worlds.
A typical language model approaches scenario generation as a writing exercise. Given a prompt about the future of artificial intelligence or climate policy, it produces a coherent narrative describing one possible trajectory. The result is often compelling prose, but beneath the surface there is little structure. Facts drift over time, causal relationships are implied rather than represented, and generating multiple scenarios usually means sampling different paragraphs instead of exploring genuinely different futures.
This is not a failure of language models. It is a consequence of what they were designed to do. They predict the next token in a sequence.
Scenario generation demands something fundamentally different.
Instead of predicting the next word, a scenario engine should predict the next world state.
From narrative continuation to state transition
Imagine describing today's economy as a structured state: unemployment, energy prices, AI capabilities, regulation, geopolitical relationships, industrial productivity, and public trust. Now introduce an intervention—a breakthrough in autonomous scientific discovery, or a coordinated international AI treaty. The objective is not to write a story about the consequences. The objective is to generate the next valid state of the world while preserving causal consistency.
Repeat this process and a future emerges—not because a model is telling a story, but because it is procedurally constructing a sequence of coherent state transitions.
This is the motivation behind what we call a Procedural World Model (PWM) for scenario generation.
Rather than treating scenarios as documents, we treat them as generated objects.
Each scenario is a structured trajectory through a high-dimensional state space. Different interventions produce different trajectories. Some converge toward similar outcomes. Others diverge dramatically. The role of the model is to learn how these trajectories evolve, not how they should be described in English.
Why procedural generation matters
This distinction matters because procedural generation unlocks capabilities that narrative generation cannot easily provide.
- Consistency. Generated scenarios become internally consistent because every transition operates on an explicit representation of the world.
- Meaningful branching. Different futures arise from different state transitions rather than different writing styles.
- Measurable evaluation. Scenarios can be compared using changes in state rather than subjective judgments about prose quality.
The long-term goal is not to build another chatbot for foresight. The goal is to develop a foundation model whose native object is the evolving world itself.
What would such a model be trained on?
If language models are trained on text, what would a Procedural World Model be trained on?
Our hypothesis is that the training data should consist of procedural transitions rather than completed narratives.
Instead of learning from pairs of prompts and essays, the model would learn from examples of structured state evolution. Every training example represents a world at time t, an intervention or set of events, and the resulting world at time t + 1. These examples could be drawn from historical data, synthetic simulations, policy exercises, strategic games, and expert-authored futures.
Over time, the model learns not merely what happened, but how worlds evolve under intervention.
Fine-tuning then becomes fundamentally different from conventional instruction tuning. We are not teaching the model to write better scenarios. We are teaching it a transition function over structured worlds.
Inference as exploring futures
Once trained, inference also changes.
A user no longer asks, "Write me three scenarios for AI in healthcare."
Instead, they specify an initial world state, define one or more interventions, choose a planning horizon, and allow the model to generate a branching space of plausible futures. Narrative summaries can still be produced, but they are downstream products. The scenario itself is the underlying procedural object.
This is why we describe the research as procedural generation for strategic foresight rather than language generation for scenario writing.
An interactive prototype: the Scenario Explorer
We built a working demonstration of this loop. The Scenario Explorer runs on a real foresight training set — thousands of examples, each pairing a world state and a policy intervention with a documented next state and an analyst's note.
Opening a scenario drops you into a walk. On one side you read the story so far: the current state of the world and the intervention being tested — say, "China agrees to Plan A but secretly and aggressively defects by stockpiling hard-to-track chips." On the other side you face a grid of possible next states. Some are documented endings from the training data; the rest are drawn by the generator.
The generator is deliberately exposed rather than hidden behind a "regenerate" button:
- seed = hash(scenario · path · slot) — the seed is derived from the scenario, the exact sequence of choices you have made, and the option's position
- pick = seed mod pool — the seed selects deterministically from the pool of transitions
Nothing is random. Replay the same choices and the same futures come back. Backtrack to an earlier step, choose differently, and everything downstream re-derives — the branching structure is the mechanism, visible on screen.
A walk is finite: six steps to the horizon. At any point you can complete the story, which keeps every choice you made, auto-derives the remaining chapters along your path, and lays the whole trajectory out as one linear read — prologue, intervention, chapters — the way a scenario narrative like AI 2027 reads, except here the narrative is assembled from explicit state transitions rather than written as free prose.
The prototype makes the thesis concrete: the scenario is the procedural object; the story is a projection of it.
The research contribution
It is tempting to frame this work as "an LLM fine-tuned for scenarios." That undersells the research problem. It makes the effort sound like another prompt-engineering project.
The interesting hypothesis is sharper: current foundation models are optimized to generate linear sequences—text, code, tokens—while scenario generation is fundamentally a procedural generation problem. A scenario is not a paragraph. It is a sequence of state transitions constrained by causality, consistency, and interventions.
That is the thesis. The names that communicate it are Procedural World Model (PWM) or Procedural Foundation Model for Scenario Generation (PFM). Both emphasize the technical novelty: the model is learning world evolution, not narrative continuation.
Current foundation models transformed software by learning to generate text and code. We believe the next generation of models for foresight will learn to generate structured futures.
The ambition is not simply to tell better stories about tomorrow.
It is to build models that can explore the space of possible tomorrows themselves.