The one-sentence version
Today's large models mainly learn "which word usually follows which word" from text. This makes them excellent at describing the world, but not necessarily at understanding how the world works: whether a cup will tip over if pushed, or what happens if conditions change. The concept proposed here is: during the post-training stage, let the LLM act within a three-dimensional or even higher-dimensional world model. The training signal is no longer "whether a human thinks this answer is good," but rather "is your predicted next world state correct" and "did your intervention yield the expected result." This shifts the training objective from correlation toward causality.
Positioning
This is a research concept that has not yet been fully implemented. This article distinguishes between three types of content: Existing Evidence (published, reproducible), Reasonable Speculation (derived from existing results), and Hypotheses to be Verified (requiring experimentation).
Layer 1: The Intuition
1.1 A person who has read ten thousand books on swimming
A person who has read every book on swimming can explain the principles of the movement perfectly. But if thrown into the water, they might not necessarily know how to swim.
Large models are somewhat like this person: they have learned the description of the world from massive amounts of text. The problem is:
- Patterns in text are mostly correlational: "rain" and "umbrella" often appear together;
- But questions like "if I close the umbrella, will the rain stop?" require causal knowledge;
- Textual descriptions of the physical world are sparse and biased: no one writes "what happens if a cup is placed 0.3 cm from the edge of a table and is bumped by an elbow."
1.2 The Concept: Giving the model a "Swimming Pool"
Let the model enter an interactive simulated world:
- The model observes the current state (object position, velocity, attributes);
- The model proposes an action and predicts what the world will look like after the action;
- The world model (or physical simulator) provides the actual next state;
- The gap between the prediction and reality becomes the training signal.
In this way, the model learns not "how people speak," but "how the world changes."
1.3 Why "Above Three Dimensions"
| Dimension | Meaning |
|---|---|
| 3D Space | Object position, occlusion, collision, gravity |
| + Time | Dynamics, causal sequence, latency effects |
| + Object Attributes | Mass, friction, elasticity, and other latent variables |
| + Abstract States | Beliefs and intentions of other agents, state of economic systems, program execution states |
"High-dimensional" does not just refer to spatial dimensions, but to the entirety of the degrees of freedom of the world state.
Layer 2: Existing Evidence
2.1 World Models are already very strong
| Work | Content |
|---|---|
| Dreamer series (Hafner et al., DreamerV3 released in 2023, published in Nature in 2025) | Learns a world model of the environment in latent space and trains policies through "imagination"; uses the same set of hyperparameters to solve many different tasks, including mining diamonds in Minecraft from scratch |
| MuZero (DeepMind, 2019–2020) | Learns a predictive model without knowing the game rules, reaching top-tier levels in Go, Chess, and Atari games |
| JEPA / V-JEPA 2 (LeCun's team, 2023–2025) | Does not predict pixels, but predicts occluded parts in the representation space; V-JEPA 2 can be used for robotic planning after pre-training on video |
| Genie 2 / Genie 3 (DeepMind, 2024–2025) | Generates interactive 3D environments from a single image or a piece of text, allowing for real-time control |
| Cosmos (NVIDIA, 2025) | A "world foundation model" platform for robotics and autonomous driving, used to generate physically plausible synthetic training data |
2.2 LLM causal reasoning indeed has flaws
- Benchmarks like CLadder (2023) test LLMs according to Pearl's Ladder of Causation, finding that model performance on intervention and counterfactual questions is significantly worse than on associative questions, and is easily disrupted by changes in phrasing;
- Extensive research shows that LLMs exhibit unstable performance on intuitive physics problems (e.g., whether an object will fall, how liquid flows).
2.3 Pearl's Ladder of Causation
| Level | Question | Requirement |
|---|---|---|
| 1. Association | What is the probability of Y given X? | Observational data (text is sufficient) |
| 2. Intervention | What happens to Y if I do X? | Ability to perform experiments or knowledge of causal structure |
| 3. Counterfactual | What would have happened to Y if I hadn't done X? | A complete causal model |
Pearl's core argument is: Based solely on Level 1 data, it is theoretically impossible to answer Level 2 and 3 questions (unless there are additional structural assumptions). Pure text pre-training mainly stays at Level 1; acting within a world model is inherently Level 2.
Layer 3: Specific Forms of the Concept
3.1 Training Loop (Reasonable Speculation)
The reward consists of two parts:
- Prediction Accuracy: The gap between the model's prediction of the next state and the actual state (measured in representation space, rather than pixel-by-pixel);
- Goal Achievement: Whether the model successfully acts according to a specified goal.
The former forces the model to learn "how the world works," while the latter forces it to learn "how to utilize this knowledge."
3.2 Comparison with existing RLHF
| RLHF | World Model Post-training | |
|---|---|---|
| Reward Source | Human or Reward Model preferences | The real response of the world |
| What is learned | What kind of answer is likable | What kind of action produces what consequence |
| Reward Hacking Risk | High (pleasing, sycophancy) | Lower (physics cannot be flattered), but simulator loopholes may still be exploited |
| Data Cost | Manual annotation is expensive | Simulators can generate infinite data |
3.3 Why not just train a world model directly?
Because language provides the ability for abstraction and composition: a world model that only understands pixels is not good at abstract reasoning like "sort these three boxes by weight." The key concept is to let the language model and the world model constrain each other: language provides abstraction, and the world provides grounding (this also responds to philosophical critiques like the Chinese Room and Symbol Grounding).
3.4 Hypotheses to be Verified
- Can causal knowledge obtained in a simulated world transfer to pure text-based causal reasoning tasks?
- Will simulator bias (simplified physics, missing phenomena) cause the model to learn incorrect "physics"?
- How many dimensions and how much realism are required? Does a "sufficiently abstract and cheap" world (e.g., simplified 2D physics plus symbolic states) provide most of the benefits?
- Will this type of training damage the model's original linguistic capabilities?
3.5 A feasible pilot experiment
- Environment: An open-source physics simulator (e.g., MuJoCo, a simplified block world);
- Interface: Use structured text to describe states; the model outputs actions and predictions of the next state;
- Control Group: A model of the same scale using only text for supervised fine-tuning;
- Evaluation: CLadder-style causal benchmarks, intuitive physics Q&A, and unseen new physical scenarios.
Summary
| Layer | Key points |
|---|---|
| Problem | Text pre-training mainly learns correlation; intervention and counterfactual reasoning are weak |
| Concept | Act within a high-dimensional world model, using "prediction of the next state" and "intervention success" as rewards |
| Basis | Pearl's Ladder of Causation; success of world models like Dreamer, MuZero, JEPA, Genie, etc. |
| To be Verified | Transferability, simulator bias, required world complexity, impact on linguistic capabilities |