Looped World Models introduce iterative latent depth as a new scaling axis for world simulation
A new arXiv preprint introduces Looped World Models (LoopWM), a parameter-shared transformer architecture that iteratively refines latent environment states to achieve up to 100x parameter efficiency over conventional world models. The approach uses adaptive computation to scale depth dynamically per prediction step, addressing the tension between long-horizon simulation fidelity and deployment cost. The authors position iterative latent depth as a new scaling axis orthogonal to model size and training data.
Related guides (4)
Agent and Tool EcosystemTopic guide
Agent and Tool Ecosystem: How the Infrastructure Layer Around LLMs Is Consolidating
Related events (8)
Looped Diffusion Language Models (LoopMDM): Depth Scaling via Layer Looping
LoopMDM introduces selective looping of early-middle transformer layers in masked diffusion language models, achieving a depth-scaling effect without adding parameters. The approach matches same-size MDM performance with up to 3.3× fewer training FLOPs and outperforms deeper non-looped MDMs on reasoning benchmarks, including up to 8.5 points improvement on GSM8K. Inference-time compute scaling is enabled by varying loop counts, with adaptive loop scheduling providing additional efficiency gains. Attention analysis suggests looping works by promoting interactions among masked token positions.
AHA-WAM: Asynchronous world-action modeling with temporal decoupling for robot manipulation
AHA-WAM introduces a dual Diffusion Transformer architecture that decouples world prediction (low-frequency) from action execution (high-frequency) in robot manipulation policies, addressing the inefficiency of existing world-action models that force both branches to operate at the same temporal resolution. The system uses a rolling key-value memory video DiT as a long-horizon scene planner and a fast action DiT that queries layerwise latent context via joint attention, with Observation-Guided Video-Context Routing enabling asynchronous execution. On RoboTwin benchmarks, AHA-WAM achieves 92.80% average success and 78.3% on real-world tasks at 24.17 Hz, a 4.59x speedup over Fast-WAM, without robot-data pretraining.
WorldKernel: Formalizing world models as coupling kernels over counterfactual worlds
A new arXiv preprint identifies a structural failure mode in prediction-based world models: strong predictors can recover the diagonal of a counterfactual coupling kernel (ordinary posteriors) but systematically fail on off-diagonal cross-world couplings, collapsing to point estimates that are sometimes provably inadmissible. The authors formalize a world model as a positive semidefinite kernel K(T,T') over admissible possible worlds, showing the off-diagonal encodes counterfactual structure that more data cannot resolve. They demonstrate that PSD constraints provide partial identification bounds computable in polynomial time, that ontological axioms tighten these bounds, and that targeted constraint learning ('scars') closes the gap faster than untargeted approaches. The work has implications for causal reasoning in AI systems and the theoretical limits of learned world models.
Latent World Recovery: multimodal learning framework for missing modalities in bioscience
A new arXiv preprint introduces Latent World Recovery (LWR), a framework for multimodal learning when some modalities are unavailable at training or inference time. LWR aligns modality-specific embeddings in a shared latent space and fuses only available modalities, avoiding explicit reconstruction of missing ones. The approach is evaluated on incomplete multi-omics benchmarks for cancer phenotype classification and survival prediction, demonstrating robustness under partial observation.
Fine-Tuning NVIDIA Cosmos Predict 2.5 with LoRA/DoRA for Robot Video Generation
This Hugging Face blog post details a workflow for fine-tuning NVIDIA's Cosmos Predict 2.5 world model using LoRA and DoRA parameter-efficient techniques for robot video generation tasks. The post covers practical implementation steps for adapting the foundation video model to robotics-specific domains. This represents a concrete application of world models to embodied AI, where synthetic video generation can support robot training data pipelines.
Variable-Width Transformers: X-shaped architecture outperforms uniform-width baselines with 22% fewer FLOPs
Researchers propose the ><former (X-shaped transformer), a decoder-only architecture that uses wider early and late layers with narrower middle layers, implemented via a parameter-free residual resizing mechanism. Evaluated on models from 200M to 2B dense parameters and 3B MoE, the architecture consistently outperforms parameter-matched uniform-width baselines on language modeling loss. The design yields a 22% reduction in FLOPs and 15% reduction in KV cache memory under fitted scaling curves, suggesting nonuniform width allocation is a viable path to more compute-efficient language models.
Walrus: A 1.3B-Parameter General Transformer for Fluid Dynamics Simulation
Polymathic AI Collaboration released Walrus, a 1.3 billion-parameter transformer model that simulates fluids, gases, and plasmas across 19 physical domains, outperforming prior specialized physics models. The model addresses aliasing artifacts in transformers—errors that compound at specific spatial locations over time—by randomly jittering input data at each time step before encoding, distributing errors rather than allowing accumulation. Walrus achieved lowest VRMSE in 18 of 19 domains for one-step predictions, reducing error by 63.6% on average versus best competing models. The jittering technique may generalize to vision and video transformer architectures where similar pixelation artifacts occur.
Training-Free Looped Transformers: Inference-Time Recurrence via ODE-Motivated Layer Reapplication
The paper introduces a method to retrofit recurrence onto frozen pretrained transformer checkpoints at inference time by looping a contiguous mid-stack block of layers without any fine-tuning or architectural changes. Naive block reapplication degrades performance, so the authors motivate their approach by treating pre-norm transformer blocks as forward Euler ODE steps and replacing one large update with smaller damped sub-steps. Evaluated across seven model families including dense, sparse MoE, and MLA+MoE architectures, the method yields consistent benchmark improvements (e.g., +2.64 pp on MMLU-Pro for Qwen3-4B-Instruct) at no training cost.