Can an AI imagine what it could do and what would happen next, instead of just guessing the next frame?
Can world models simulate actionable possibilities instead of just predicting next states?
This explores whether a world model can be built to imagine what an agent could do and what would follow (the what-ifs), instead of only forecasting what the next observation will look like.
This explores whether a world model can imagine what an agent could do and what would follow, instead of only forecasting the next observation. The corpus says yes, and treats it as the line between a real world model and a good predictor. One note points out that LLMs can score well on prediction with task-specific shortcuts, without any coherent model of how the world works. The real test is whether the model can reason about interventions and what-if scenarios, not just surface regularities What makes a world model actually useful for reasoning?. Another borrows the psychology of hypothetical thinking. It argues the thing to simulate is the whole space of actionable possibilities: physical, embodied, emotional, social, mental, counterfactual, even evolutionary. Passive frame prediction is the wrong target What should a world model actually be designed to do?.
A more engineering-minded version of the same idea asks what an agent needs back from a world model. Instead of a predicted physical state, the model returns things an agent can use: execution outcomes, retrieved past experiences, skills, and verification signals, grouped into six forms (dynamics, spatial, execution, memory, skill, reward) Should world models predict states or return agent feedback?. This fits a separate finding on very long tasks. What predicted success there was persistence, meaning repeated cycles of try, measure, and fold the result back in, not the quality of the first attempt What predicts success in ultra-long-horizon agent tasks?. A model that answers 'what happens if I do this' feeds that loop directly.
The wrinkle is that next-state prediction hasn't been dropped. A language world model trained by next-state prediction on more than 10 million agent trajectories beat training in real environments on three benchmarks and transferred across seven domains Can language models learn to simulate agent environments?. Those trajectories include the agent's actions, so the model learns what follows from a choice. The dispute looks less like prediction versus possibility and more like what the model is conditioned on and what it gets used for. Training also seems to matter: post-trained models behave as if they recognize their outputs as actions that shape their next inputs, with 3-4x lower output entropy on-policy. Pretraining alone doesn't produce that action-and-consequence loop Do models recognize their own outputs as actions shaping future inputs?.
Simulating possibilities also means simulating people. Tracking only the physical scene gives wrong predictions of human action even when the scene is right. Beliefs, wants, and intentions have to be explicit state, coupled to the physics Can world models predict human action from physics alone?. And LLMs that look socially competent when one model plays every role fail when agents hold private information. The what-if space includes what each party can't see, and models skip that grounding work when they're allowed to know everything Why do LLMs fail when simulating agents with private information?.
As for how far this has gone, the aspiration is well ahead of the evidence. In economic simulation, most work sits on the lower rungs of a six-level ladder with fixed or adaptive agents. Self-evolving agents, institutions that emerge from within the simulation, and validation against real data are rare Where do economic world models currently stand in capability?. Other work makes prediction cheaper, for example looped computation that spends more depth on harder steps Can looped computation replace parameter count in world models?. Nothing in these notes tests a possibility-simulating world model head to head against a next-state one.
Sources 10 notes
Research shows LLMs may achieve high prediction accuracy through task-specific heuristics without developing coherent generative models of how the world works. True world models must enable reasoning about interventions and counterfactuals, not surface regularities.
Drawing on hypothetical thinking in psychology, world models are most useful when designed to simulate all actionable possibility spaces—physical, embodied, emotional, social, mental, counterfactual, and evolutionary—grounded in agent decision-making rather than passive prediction.
Rather than predicting physical states, world models should function as Agent-Centric Interactive World Proxies that return execution outcomes, retrieved experiences, skills, and verification signals organized into six functional forms: dynamics, spatial, execution, memory/experience, skill, and reward/verification.
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Qwen-AgentWorld demonstrates that native language world models trained via next-state prediction on 10M+ trajectories outperform real-environment training on three benchmarks and transfer across seven domains, positioning next-state prediction as a foundation objective for agents.
Show all 10 sources
Post-trained language models exhibit a measurable shift where they recognize their outputs become their own future inputs, closing an action-perception loop absent in pretraining. Evidence includes 3-4x lower output entropy on-policy and behavioral signatures of trajectory recognition.
Research across eight LLM-based world models shows that tracking only the physical scene leads to wrong action predictions even when the scene looks correct. Mental World Modeling makes beliefs, wants, and intentions explicit state components coupled to physical simulation, and all three elements are required for accurate human decision prediction.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
A six-level ladder ranks economic world models by how much economy behavior emerges from within the simulation. Current work concentrates in lower levels with fixed or adaptive agents, while self-evolving agents, endogenous institutions, and empirical validation remain rare.
LoopWM achieves up to 100x parameter efficiency by refining latent environment states through iterative computation in a shared block, with spectral-norm constraints providing formal stability guarantees. The approach mirrors physical system recurrence, spending more depth on harder prediction steps.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Qwen-AgentWorld: Language World Models for General Agents
- Can Language Models Serve as Text-Based World Simulators?
- Looped World Models
- Mental World Modeling
- Critiques of World Models
- Is this the real life? Is this just fantasy? The Misleading Success of Simulating Social Interactions With LLMs
- Agent Learning via Early Experience
- Quo Vadis, World Modeling?