Looped World Models (LoopWM)
Hongyuan Adam Lu, Z. L. Victor Wei, Qun Zhang, Jinrui Zeng, Bowen Cao, Lingwei Meng, Mocheng Li, Zezhong Wang, Haonan Yin, Naifu Xue, Minyu Chen, Cenyuan Zhang, Zefan Zhang, Hao Wei, Jiawei Zhou, Haoran Xu, Hao Yang, Ronglai Zuo, Tongda Xu, Yonghao Li, Jian Chen, Hebin Wang, Zeyu Gao, Yang Li, Wei Zhao, Qimin Zhong, Siqi Liu, Yumeng Zhang, Leyan Cui, Zhangyu Wang, Wai Lam
Looped World Models (LoopWM) use parameter-shared transformers and deferred decoding to achieve stable, adaptive-depth environment simulation.
How can we achieve faithful long-horizon world modeling without the prohibitive compute costs and compounding errors of deep, monolithic architectures?
World models struggle with a fundamental tension: faithful long-horizon simulation requires deep computation, but fixed-depth architectures are expensive and prone to compounding errors as trajectory quality degrades over time. LoopWM resolves this by replacing fixed-depth backbones with a parameter-shared transformer block that iteratively refines latent states, using spectral constraints to ensure stable rollouts across arbitrary horizons. This architecture achieves up to 100× parameter efficiency compared to conventional models, while supporting adaptive computation that scales depth based on the complexity of each transition.
Paper Primer
The core mechanism is a recurrent dynamics core that applies a shared transformer block $T$ times to a latent state, effectively decoupling computational depth from parameter count. The model uses a spectrally-constrained state-retention matrix to guarantee that latent updates remain bounded, preventing the numerical explosions common in deep recurrent models.
To further improve efficiency, the authors introduce "deferred decoding," which eliminates intermediate observation reconstruction during multi-step rollouts. The model performs latent reasoning across a sequence of actions and decodes only at the final step: think first, decode once.
LoopWM achieves superior predictive accuracy and stability while using significantly fewer parameters than fixed-depth baselines.
On the ScienceWorld benchmark, the 1B-parameter LoopWM outperforms the much larger Claude-Opus-4-6-max, improving average EM scores by 21.2%. Up to 100× parameter efficiency over conventional world models.
Adaptive computation depth allows the model to allocate resources dynamically based on transition difficulty.
Simple transitions (e.g., free flight) exit the loop after minimal iterations, while complex events (e.g., collisions) trigger deeper refinement, reducing aggregate inference FLOPs by up to two orders of magnitude.
Why is a looped architecture better suited for world modeling than a standard, fixed-depth transformer?
Physical dynamics are inherently iterative, involving the repeated application of stationary laws; a looped transformer structurally mirrors this process, allowing the model to refine its latent state estimate through repeated passes rather than relying on a single, rigid forward pass.
What is the primary benefit of "deferred decoding" in this architecture?
It prevents the model from wasting representational capacity on pixel-level reconstruction at every intermediate step, allowing the latent trajectory to focus on action-relevant structure over extended horizons.
LoopWM establishes "iterative latent depth" as a new scaling axis for world simulation, orthogonal to traditional scaling via model size or training data volume.
Researchers can now treat computational depth as a flexible, test-time variable rather than a fixed architectural constraint, enabling high-fidelity simulation on resource-constrained hardware.
The Tension in World Modeling
We expose the compute‑error trade‑off in world models and introduce LoopWM to resolve it.
World models must predict long‑horizon dynamics, yet faithful simulation traditionally demands deep computation, which inflates parameter count and amplifies error accumulation across rollout steps.
LoopWM tackles this by applying a single transformer block repeatedly to a latent state, sharing parameters across iterations and iteratively refining the prediction; the authors report up to 100× parameter efficiency compared to conventional fixed‑depth models and an adaptive mechanism that scales loop depth to match the difficulty of each transition.
The architecture employs a spectrally constrained state‑retention parameterisation (e.g., $A = \operatorname{diag}(-\exp(a))$) that keeps the recurrent update contractive, guaranteeing bounded dynamics for arbitrarily long rollouts, while the adaptive‑depth property allocates more loop iterations to complex events (collisions) and fewer to simple ones (free flight).
The image displays a stylized infinity symbol logo with a gradient color transition from blue on the left to purple on the right.
The core trade‑off is between model depth (or loop iterations) and simulation stability: deeper iterative refinement improves fidelity but must be balanced against computational cost.
Foundations of World Modeling
We survey prior world‑model and looped‑transformer approaches, highlighting their strengths and gaps.
This section situates our Looped World Model within the broader landscape of world‑model research and looped‑transformer designs.
Latent dynamics model (RSSM) that plans directly in latent space from pixel observations.
Model‑based Atari agent that learns a video‑prediction simulator as a learned environment.
Combines a learned dynamics model with Monte‑Carlo tree search to master board games and Atari without ground‑truth rules.
RSSM‑based world models that achieve human‑level performance across >150 tasks with a single hyper‑parameter set.
Replaced the recurrent backbone of Dreamer‑style models with an autoregressive transformer over discrete tokens.
Introduced a Transformer State‑Space Model for memory‑intensive world‑model tasks.
Improved tokenization efficiency via context‑aware delta encoding.
Used diffusion models to produce visually faithful world simulations.
State‑of‑the‑art Crafter performance using masked generative transformers over spatial latent states.
Long‑form video generation with emergent 3D consistency, cast as a world simulator.
Text‑conditioned generative model that produces interactive, explorable environments.
Extended Genie with higher resolution and richer scene dynamics.
Comprehensive overviews of world‑model applications in autonomous driving, embodied AI, and video generation.
Introduced weight sharing across transformer layers combined with Adaptive Computation Time.
Demonstrated full cross‑layer parameter sharing in BERT‑scale models.
Theoretical proof that looped transformers can simulate arbitrary programs with constant parameters.
Showed looped transformers match standard transformer performance on in‑context learning while using <10% of the parameters.
Demonstrated length‑generalisation improvements via adaptive loop counts.
Provided empirical evidence that looped models generate “latent thoughts,” enabling deeper reasoning.
Trained looped language models (LoopLMs) through the full LLM pipeline, achieving 2–3× parameter efficiency.
Introduced recurrent‑depth models that scale test‑time compute by increasing loop count.
Analyzed geometry of latent dynamics in recurrent‑depth transformers, revealing fast intra‑loop and slow inter‑token dynamics.
Elastic‑depth training with shortcut modulation for budget‑conditioned reasoning.
Per‑token dynamic depth allocation within a single recursive framework.
Combined mixture‑of‑experts with universal transformers to balance specialization and sharing.
Allows recurrent networks to halt computation per step based on a learned halting probability.
Techniques that terminate inference at intermediate layers when confidence is sufficient.
Tokens exit the loop when prediction entropy falls below a learned threshold.
Samples Poisson‑distributed loop counts during training to induce robustness to variable test‑time depth.
Conditions training on a continuous “time budget” to enable fine‑grained compute allocation.
Uses second‑order differences of hidden states to decide when to halt loops.
Despite these advances, none of the prior works combine adaptive‑depth looping with world‑model simulation, leaving the compute‑vs‑error tension unresolved.
The LoopWM Architecture
LoopWM iteratively refines latent states with a spectrally‑stable shared transformer and defers decoding to the final step.
World‑model accuracy degrades when the latent dynamics require many sequential transformer layers, yet each extra layer multiplies compute and amplifies error accumulation.
LoopWM treats each physical transition as a short, iterative refinement inside a shared transformer loop, reusing the same parameters while the number of refinement steps adapts to transition difficulty.
Iteration 0: $h^{(1)} = \bar{A}h^{(0)} + \bar{B}e = [0,0] + [1,-1] = [1,-1]$.
Iteration 1: $h^{(2)} = \bar{A}h^{(1)} + \bar{B}e = [0.5, -0.5] + [1,-1] = [1.5,-1.5]$.
The diagonal $\bar{A}$ halves the previous hidden state each iteration, so the hidden trajectory stays bounded regardless of how many loops are executed.
How does LoopWM differ from a standard recurrent neural network that also applies the same cell repeatedly?
LoopWM’s recurrent block is a full transformer layer (self‑attention + feed‑forward) with a spectrally‑constrained linear skip connection $\bar{A}$. An RNN typically uses a simple affine update without attention and lacks the explicit $\rho(\bar{A})<1$ guarantee, so it can explode or vanish over many steps.
Training a fixed‑depth transformer for every transition is wasteful because many dynamics are simple; LoopWM therefore samples a loop count $T$ from a Poisson distribution, letting easy steps use few iterations and hard steps use more.
Even with adaptive loop counts, many transitions converge after a single refinement. LoopWM therefore equips the recurrent block with a lightweight exit gate that can halt the inner loop early, saving compute on trivial steps.
The model watches its own hidden state each refinement step and stops looping once a learned gate signals that further updates would be negligible.
Why not simply reduce the number of transformer layers globally instead of using an early‑exit gate?
A global reduction would force every transition to share the same coarse approximation, harming accuracy on complex dynamics. The gate lets easy transitions stop early while still allowing hard transitions to consume the full $T$ iterations, preserving fidelity where it matters.
Even with adaptive depth, decoding after every latent step still forces the model to allocate capacity to reconstruct observations that are never used for planning. Deferred Decoding removes all intermediate decodings, letting the latent dynamics focus on long‑term action‑conditioned reasoning.
Instead of decoding observations, rewards, and continuation signals after each latent transition, the model rolls forward $K$ actions entirely in latent space and decodes only once at the final step.
Step 0: $h_1 = \bar{A}h_0 + \bar{B}u_0 = [0,0] + [1,0] = [1,0]$.
Step 1: $h_2 = \bar{A}h_1 + \bar{B}u_1 = 0.5\![1,0] + [0,1] = [0.5,1]$.
Final decode: $D_\xi([0.5,1])$ produces a single observation, reward, and continuation prediction for the whole two‑step plan.
By deferring decoding, the model never spends compute on intermediate reconstructions; the hidden trajectory directly encodes the effect of both actions.
How does Deferred Decoding differ from Dreamer’s imagination rollouts, which also avoid re‑encoding observations?
Dreamer still applies reward and value heads at every imagined step, incurring $K$ decoder passes. Deferred Decoding removes *all* per‑step decoder calls, invoking $D_\xi$ only once after the full $K$‑step latent rollout, which yields larger compute savings and forces the latent dynamics to capture longer‑term structure.
**Figure 1.** The overall framework of our proposed Looped World Models (LoopWM)
Table 1 contrasts LoopWM‑DD with prior world‑model families, emphasizing that only LoopWM‑DD combines looped depth, action‑conditioned latent dynamics, and deferred decoding.
Empirical Performance
Looped World Model delivers strong gains on benchmark tasks while staying tiny.
The LoopWM architecture trades deep sequential computation for an iterative transformer loop, preserving long‑horizon fidelity while keeping the parameter count low.
A suite of simulated physics and biology tasks where a model must predict the outcome of a sequence of five actions.
A large closed‑source language model (≈100 B parameters) released by Anthropic, used as a strong commercial baseline for world‑modeling.
A compact open‑source model (≈6 B parameters) optimized for fast inference, serving as a lightweight baseline.
Looped World Model attains 68.4 % exact‑match on ScienceWorld, a 21.2 % absolute gain over Claude‑Opus‑4‑6‑max.
Table 2 shows EM 68.4 % for LoopWM versus 47.2 % for the baseline.
Claude‑Opus‑4‑6‑max reaches 47.2 % EM while running on a model that is more than 100 × larger (≈100 B parameters). LoopWM’s 1 B‑parameter footprint makes it far cheaper to deploy.
**Figure 2.** Relative increase over Qwen3.7-max on automatic online performance, compared against baselines. Note that the results are obtained via online estimation, with the tasks of danmaku generation. LWM denotes LoopWM.
On the AlfWorld suite (Table 4), LoopWM secures the highest BLEU score among four models and ranks second on EM and Token F1, despite its modest 1 B‑parameter size.
Entity extraction remains the weakest dimension for LoopWM on AlfWorld, suggesting future work should target richer object‑level representations.
Across all tables, deferred decoding consistently improves performance, especially when rollouts are accumulated, confirming its utility for long‑horizon planning.
Summary and Future Directions
Concluding the work, we highlight LoopWM’s efficiency, stability, and new scaling axis.
Looped World Model (LoopWM) applies a looped transformer architecture to world modeling, iteratively refining latent states with a shared transformer block and stabilised residual dynamics. This yields up to 100× parameter efficiency while guaranteeing stable rollouts via spectral‑norm constraints and adapting computation depth to the difficulty of each transition. We propose iterative latent depth as a new scaling axis orthogonal to model size and data volume, opening paths to more capable, efficient, and stable long‑horizon simulations.
Broader Impacts
Broader impacts and scope of the LoopWM design are discussed.
This manuscript deliberately limits its scope, focusing on establishing the core architectural thesis—looped latent refinement, deferred decoding, and stabilized dynamics—as a viable design space for world modeling, rather than exhaustively reporting every supporting result.
The presented experiments already illustrate the benefit of iterative latent computation via deferred decoding, showing that preserving and refining latent states across rollout steps improves performance.
Table 4 compares the LoopWM (≈ 1 B parameters) against Claude‑Opus‑4‑6‑max and other baselines on the AlfWorld benchmark, reporting metrics such as EM, F1, BLEU, and entity accuracy across multiple task variants.
Aggregate Performance Metrics
LoopWM delivers top‑line Token F1 of 83.5 % while improving EM and BLEU over strong baselines.
LoopWM attains a Token F1 of 83.5 % on the benchmark, outpacing all baselines.
Claude Opus 80.4 %, Qwen‑3.5‑flash 72.6 %, Gemini‑3‑flash‑preview 67.3 %.
Across the five test conditions (clean, heat, pick, look, cool) LoopWM consistently raises Token F1, reaching 84–86 % on the hardest “heat” and “pick” settings, whereas the baselines dip below 80 %.
The authors note that looping directly contributes to these gains: iterative transformer updates refine predictions without inflating parameter count.
Beyond the reported domains, the architecture also succeeds in continuous visual environments, where training loss steadily decreases, confirming its broader applicability.
A current limitation is that the manuscript does not yet disclose the full breadth of validation experiments, leaving some empirical evidence unpublished.
Positionally, LoopWM occupies a distinct point in the world‑model landscape, differing from RSSM‑style latent dynamics, autoregressive video token models, and diffusion‑based approaches.
Ablation: Deferred Decoding Effects
Deferred decoding’s step‑wise gains are quantified across all ScienceWorld tasks.
This ablation isolates the contribution of Deferred Decoding by comparing LoopWM with and without the component across five incremental steps on each ScienceWorld sub‑task.
Deferred decoding yields a +700.9 % Entity boost on the Boil task at Step 4.
Table 6 reports the largest single‑task improvement (+700.9 % Entity) among all metrics.
Improvements grow with later steps, peaking at +585.3 % Entity on the Melt task at Step 4.
Table 14 shows the highest Entity gain (+585.3 %) and a consistent upward trend across steps.
The table presents performance improvements across five steps (Step 1 to Step 5) for four metrics: EM, F1, BLEU, and Entity.
**Table 6.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Boil, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines) / Baselines * 100\%$. Note that our model is a model with about 1B model parameters. '—' represents that the baseline score is 0 and LoopWM score is not zero. '0%' means that both of them has a score of 0.
**Table 7.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Chemistry, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance (Ours - Baselines)/Baselines * 100%. Note that our model is a model with about 1B model parameters.
**Table 8.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Conductivity, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines) / Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 9.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Find, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines) / Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 10.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Freeze, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines)/Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 11.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Genetics, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance (Ours - Baselines)/Baselines * 100%. Note that our model is a model with about 1B model parameters.
**Table 12.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Grow, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines)/Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 13.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Incline, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines)/Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 14.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Melt, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance $(Ours - Baselines)/Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
**Table 15.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task on the task of Power, compared to gemini-3-flash-preview-thinking. The relative improvements are calculated using the absolute performance (Ours - Baselines)/Baselines * 100%. Note that our model is a model with about 1B model parameters.
**Table 16.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling task averaged on all tasks, compared to qwen3.5-flash. The relative improvements are calculated using the absolute performance $(Ours - Baselines) / Baselines * 100\%$. Note that our model is a model with about 1B model parameters.
These results substantiate the claim that iterative latent depth—implemented via Deferred Decoding—acts as a natural scaling axis, with later steps delivering larger relative gains across diverse world‑modeling tasks.
Ablation: Decoding Across Tasks
Deferred decoding boosts performance on ScienceWorld tasks, delivering up to several‑thousand‑percent gains on specific metrics.
Deferred decoding improves exact‑match (EM) by +143.5 % on average across all tasks.
Table 17 reports a +143.5 % EM gain.
Deferred decoding raises BLEU by +2428.1 % on the Chemistry sub‑task (Step 2).
Table 18 lists a +2428.1 % BLEU increase at Step 2.
Deferred decoding boosts EM by +600.9 % on the Conductivity sub‑task (Step 3).
Table 19 shows a +600.9 % EM improvement at Step 3.
Deferred decoding lifts EM by +699.2 % on the Find sub‑task (Step 1).
Table 20 records a +699.2 % EM gain at Step 1.
Deferred decoding increases BLEU by +750.0 % on the Freeze sub‑task (Step 3).
Table 21 reports a +750.0 % BLEU improvement at Step 3.
Deferred decoding improves EM by +203.5 % on the Genetics sub‑task (Step 1).
Table 22 lists a +203.5 % EM gain at Step 1.
Deferred decoding raises EM by +450.4 % on the Grow sub‑task (Step 1).
Table 23 shows a +450.4 % EM improvement at Step 1.
Deferred decoding yields a modest EM gain of +40.2 % on the Incline sub‑task (Step 1).
Table 24 records a +40.2 % EM increase at Step 1.
Deferred decoding boosts EM by +822.8 % on the Melt sub‑task (Step 5).
Table 25 reports a +822.8 % EM gain at Step 5.
Deferred decoding raises BLEU by +891.7 % on the Power sub‑task (Step 4).
Table 26 lists a +891.7 % BLEU improvement at Step 4.
Deferred decoding achieves an F1 score of 87.5 % at Step 3, the highest among all steps.
Table 27 shows F1 = 87.5 % at Step 3.
Ablation: Step-wise Performance
Deferred decoding’s impact across steps and tasks is quantified in Tables 28‑36.
This ablation isolates the contribution of deferred decoding by reporting task‑level metrics for each incremental step of the Loop‑WM pipeline.
**Table 28.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 1, on our model. Note that our model is a model with about 1B model parameters.
**Table 29.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 2, on our model. Note that our model is a model with about 1B model parameters.
**Table 30.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 3, on our model. Note that our model is a model with about 1B model parameters.
**Table 31.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 4, on our model. Note that our model is a model with about 1B model parameters.
The table presents performance metrics across various task types, including EM, F1, BLEU, and Entity scores.
**Table 33.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on gemini.
**Table 34.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 1 on gemini.
**Table 35.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 2 on gemini.
**Table 36.** The effect of deferred decoding on the ScienceWorld dataset (Wang et al., 2022) world modelling tasks, on Step 3 on gemini.
Ablation: Final Performance Metrics
Ablation tables reveal how Deferred Decoding shapes task performance across steps and models.
Deferred Decoding is the ablated component; each table reports the resulting EM, F1, BLEU, and Entity scores when the component is removed.
Deferred Decoding lifts Exact Match on the “Incline” task from 0 % at Step 1 to 59.3 % by Step 4.
Table 40 (Step 1) shows 0 % EM; Table 43 (Step 4) shows 59.3 % EM for Incline.
On the “Boil” task, EM stays at 0 % through Steps 1‑4 and only reaches 11.1 % at Step 5.