Should an AI's mental model of the world get smarter by looping on one idea, or by stacking several layers of abstraction?
Should world models scale by looping depth or stacking distinct representations?
This explores two ways to make world models (AI systems that predict how an environment will change) more capable: run the same computation block several times, or build a hierarchy of separate representations where each level views the world at a different level of abstraction.
This question sets two scaling strategies for world models against each other. The first is looping: refine a single internal picture of the world by passing it through the same layers again and again. The second is stacking: give the model several distinct representations, each more abstract than the one below it. The corpus has strong evidence for both. It has no paper that tests them head to head, though, and the evidence suggests they solve different problems, so 'which one' may be the wrong question.
The argument for looping is efficiency and adaptability. LoopWM refines its estimate of the environment's state by running one shared block repeatedly. It reaches up to 100x parameter efficiency and can spend more passes on harder prediction steps, which the authors compare to the way physical systems evolve step by step through time Can looped computation replace parameter count in world models?. The same pattern shows up outside world models. Looped models beat larger feedforward networks on reasoning because repeating a computation helps them keep track of state as it changes Can models learn by looping instead of growing larger?. Selectively looping a few early-middle layers in diffusion language models beats simply adding layers, at roughly a third of the training compute Can looping layers beat adding depth in diffusion models?. A looped model can also stay interpretable: LOTUS supervises each loop step with written-out reasoning and matches explicit chain-of-thought on math Can latent reasoning close the scaling gap with explicit chain-of-thought?.
The argument for stacking is different. Some predictions need a different vocabulary, not more refinement. H-JEPA raised maze-navigation planning success from 18% to 73%. The gain came from giving each level of the hierarchy its own, more abstract latent space instead of sharing one space across levels. Higher levels could then judge possible futures in terms closer to the goal, not in terms of raw visual detail Does each hierarchy level need its own latent space?. A looped block keeps refining one kind of representation, and this result suggests that alone may not get a model to the abstraction it needs. A related finding comes from models that predict human behavior. World models that track only the physical scene predict people's actions wrongly. They need beliefs, wants, and intentions as separate, linked parts of their state Can world models predict human action from physics alone?. That is another case where separate representations matter more than additional passes.
Depth itself is also more than a dial for gradual gains. In reinforcement learning, increasing network depth produces sudden new behaviors at specific thresholds: walking appears at 16 layers and wall-climbing at 256 Does network depth unlock qualitatively new behaviors in RL?. Small language models also gain more from going deep and thin than from going wide Does depth matter more than width for tiny language models?. Neither result says whether those layers should be repeated copies or distinct ones. Reasoning research adds a caution: spending compute on depth alone can trap a model in one line of thought, and abstractions help it explore alternatives first Can abstractions guide exploration better than depth alone?.
The likely answer is that both strategies belong in the same model, used for different jobs. Looping fits steps where the world changes in a consistent way that iterative refinement can capture. Separate levels fit the move between levels of meaning: pixels to places, places to goals, bodies to minds. That points to a test the corpus hasn't run yet: a hierarchy where each level loops inside its own space. It also connects to a broader concern in the collection, that a world model is only useful if it can reason about interventions and counterfactuals, not just predict the next observation What makes a world model actually useful for reasoning?. Scaling either axis is only worth it if it moves the model toward that.
Sources 10 notes
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.
Models that re-apply layers in recurrent depth outperform larger feedforward networks on reasoning tasks. This works because recursion enables state tracking and compositional generalization that parameter scaling alone cannot achieve, with convergence signals providing natural halting.
LoopMDM matches same-size masked diffusion models with 3.3× fewer training FLOPs and exceeds deeper non-looped baselines on reasoning tasks. Reusing computation through selective early-middle layer loops proves more effective than adding depth at fixed parameter budgets.
LOTUS, a looped Transformer trained with gold CoT supervision at each latent position, closes the latent-to-explicit gap on GSM8K and cuts thought-phase latency by 2.5× to 6.9×. Latent reasoning need not be verbalized token-by-token to remain legible and competitive.
H-JEPA improves Visual AntMaze planning from 18% to 73% success by giving each hierarchy level a distinct, more abstract latent space rather than sharing one. This per-level abstraction lets higher levels score candidate futures in abstract space better matched to goal-like objectives.
Show all 10 sources
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.
Scaling to 1000-layer networks in self-supervised RL produces dramatic capability jumps at specific thresholds—depth 16 enables walking, depth 256 enables wall-climbing—driven by synergistic gains in both exploration and expressivity rather than gradual improvement.
MobileLLM shows deep-and-thin architectures yield 2.7–4.3% accuracy gains over balanced designs at 125M–350M scale by composing abstract concepts through layers rather than spreading parameters across width.
RLAD jointly trains abstraction and solution generators, showing that allocating test-time compute to diverse abstractions outperforms parallel solution sampling at large budgets. Abstractions create structured breadth-first exploration that prevents the underthinking failure mode of depth-only reasoning chains.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
- Looped World Models
- Loop the Loopies!
- Hierarchical Reasoning Model
- A Mechanistic Analysis of Looped Reasoning Language Models
- Mental World Modeling
- Qwen-AgentWorld: Language World Models for General Agents
- Can Language Models Serve as Text-Based World Simulators?