OneWorld: Unified Multi-Domain Modeling

• OneWorld is a unified modeling paradigm that replaces fragmented, task-specific representations with a single shared latent substrate, enabling cross-domain inference.
• It employs innovative methodologies such as class-agnostic proposals, in-context transformer adaptation, and unified token spaces to bridge diverse applications.
• Empirical evaluations across open-world detection, world modeling, atmospheric forecasting, and 3D generation highlight both strong performance and domain-specific limitations.

Searching arXiv for the cited “OneWorld” papers and closely related work to ground the article. OneWorld is a label applied, in current arXiv usage, to several technically distinct but conceptually related research programs that replace fragmented task-specific representations with a single world-centric substrate. In the materials considered here, the term encompasses open-world detection built around class-agnostic proposals and decoupled one-stage localization (Konan et al., 2022), a transformer world model trained only on synthetic environments and adapted in-context to new dynamics (Ferreira et al., 2024), an observation-native atmospheric model that learns directly from raw ungridded Earth observations (Gong et al., 7 May 2026), and a 3D scene generator that performs diffusion in a unified 3D representation rather than in 2D image or video latents (Gao et al., 17 Mar 2026). Across these usages, the recurring objective is to learn a latent state, proposal space, or token manifold from which reconstruction, prediction, and downstream inference can be performed with minimal task-specific recoding.

1. Semantic scope and unifying idea

The term spans at least four research settings. In open-world detection, the emphasis is not on a latent world simulator but on Open-World Proposals (OWP): class-agnostic object proposals that generalize to unseen categories. In transformer world modeling, the emphasis is on a single model trained once on a synthetic prior and then adapted to a new environment without gradient updates. In atmospheric modeling, the emphasis is on an observation-native latent field that reconstructs and forecasts the Earth system directly from heterogeneous sensor measurements. In 3D scene generation, the emphasis is on a unified 3D token space that jointly encodes geometry, appearance, and semantics.

Usage of “OneWorld”	Unified substrate	Primary objective
Open-world detection	Class-agnostic proposal/objectness space	Localize seen and unseen objects
One-shot world modeling	Transformer-conditioned transition context	Simulate new environments from context
Atmospheric world modeling	Latent Observation Space (LatentOS)	Reconstruct and forecast from raw observations
3D scene generation	3D Unified Representation Autoencoder latent	Generate cross-view-consistent 3D scenes

A plausible implication is that “OneWorld” functions less as a single standardized architecture than as a research orientation toward single-state, multi-task inference. The common technical move is to avoid repeated projection into narrow intermediate spaces: background classes in closed-set detection, per-environment retraining in model-based RL, fixed forecast grids in NWP, or 2D VAE latents in 3D generation. Each system instead promotes a shared representation from which multiple outputs can be derived.

2. Open-world detection and class-agnostic proposal spaces

In the detection setting, OneWorld refers to an open-world regime in which a detector must localize object-like regions from both base classes $C^B$ and novel classes $C^N$, even though only $C^B$ is labeled during training. The central subproblem is Open-World Proposals, defined as class-agnostic object proposals intended to generalize to unseen categories. The one-stage formulation is built on FCOS, which predicts at each location $(x,y)$ class logits $\hat{K}_{x,y}$, box offsets $\hat{R}_{x,y}=[l,r,t,b]$, and centerness $\hat{C}_{x,y}$, with centerness target

$$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$

The work argues that FCOS is attractive for OWP because classification and regression/objectness are decoupled heads, making localization less vulnerable to specialization on seen categories (Konan et al., 2022).

The principal modifications are threefold. First, centerness is replaced or supplemented by an IoU prediction branch whose target is the overlap between predicted and target offsets,

$$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$

A conditional IoU head concatenates FPN features with regression outputs so that objectness depends directly on predicted box quality. Second, IOU Sampling (IS) augments standard center sampling by setting IoU targets below $0.3$ to zero, injecting hard negatives and flattening the objectness target distribution. Third, Unknown Object Masking excludes background pixels with objectness above a threshold $C^N$0 from the classification loss, reducing contamination from unlabeled objects. The classification score at inference becomes

$C^N$1

Empirically, the one-stage design narrows the gap to two-stage RPN systems while preserving classification more effectively. On COCO novel classes, centerness-only objectness yields $C^N$2 AR100, IoU objectness $C^N$3, IOU-CS-IS $C^N$4, and Conditional IOU-CS-IS $C^N$5, compared with $C^N$6 for OLN 2-RPN. When classification is added, the FCOS-based model drops by roughly $C^N$7 to $C^N$8 AR points, whereas OLN 2-RPN drops by $C^N$9. On LVIS, IOU-CS-IS improves novel AR@300 from $C^B$0 to $C^B$1, while Unknown Object Masking does not produce clear gains. The method localizes unknowns but does not perform explicit unknown-class recognition or incremental label expansion.

3. One-shot world models from synthetic priors

In model-based RL, OneWorld denotes a transformer-based One-Shot World Model (OSWM) trained exclusively on synthetic environments sampled from a prior distribution $C^B$2. The model receives transition context from a new real environment at inference time and adapts purely in-context, without gradient descent. States and actions are organized as

$C^B$3

with targets

$C^B$4

To support multiple environments with differing dimensionalities, the formulation pads state and action vectors to global maxima $C^B$5 and $C^B$6, making environment identity implicit in the transition statistics rather than explicit metadata (Ferreira et al., 2024).

The synthetic prior has two components. The NN prior assigns each state dimension its own randomly initialized MLP, so that

$C^B$7

with random hidden widths, random activations chosen from $C^B$8, residual structure, and random output scaling and shifting. The momentum prior introduces simple physical-style dynamics via position–velocity updates

$C^B$9

Rewards are produced by a randomly initialized MLP $(x,y)$0, except that with probability $(x,y)$1 reward is set to the constant value $(x,y)$2. Training follows a Prior-Fitted Networks style objective: sample a synthetic trajectory $(x,y)$3, choose a context cutoff $(x,y)$4, and predict the future targets $(x,y)$5 with MSE loss,

$(x,y)$6

At inference, OSWM is given roughly $(x,y)$7 transitions from a new environment, normalizes the context, and predicts future next-state and reward pairs for imagined rollouts. PPO is then trained entirely inside this learned simulator. The reported performance is environment-dependent. On GridWorld, OSWM-PPO achieves $(x,y)$8, matching PPO on the real environment and far exceeding random at $(x,y)$9. On CartPole-v0, OSWM-PPO reaches $\hat{K}_{x,y}$0, satisfying the solved criterion of average reward $\hat{K}_{x,y}$1. On SimpleEnv, OSWM-PPO obtains $\hat{K}_{x,y}$2, near PPO’s $\hat{K}_{x,y}$3 and much better than random at $\hat{K}_{x,y}$4. Transfer to more complex environments remains limited: MountainCar-v0 remains at $\hat{K}_{x,y}$5, Pendulum-v1 at $\hat{K}_{x,y}$6, and Reacher-v4 shows only partial success at $\hat{K}_{x,y}$7. The paper attributes these failures to prior mismatch, insufficient context coverage, and the tendency of OSWM to learn smoother reward landscapes than the real environment.

4. Observation-native atmospheric world modeling

Earth-o1 generalizes the OneWorld idea to geophysical modeling by defining an observation-native atmospheric world model trained directly on raw, ungridded observations rather than on reanalysis fields. Its central representation is the Latent Observation Space (LatentOS), a continuous latent manifold representing the 3D atmospheric state and closely coupled surface or polar variables at arbitrary locations and times. Instead of imposing a fixed forecast grid and then assimilating observations into it, Earth-o1 tokenizes each observation in native geometry, fuses heterogeneous tokens in a shared multimodal latent space, and learns temporal evolution as latent dynamics. In this formulation, local physical variables are reconstructed through decoders of the form

$\hat{K}_{x,y}$8

where $\hat{K}_{x,y}$9 is the latent token set at time $\hat{R}_{x,y}=[l,r,t,b]$0 (Gong et al., 7 May 2026).

The system ingests LEO satellite swaths, GEO imagery, and in-situ streams such as surface stations, radiosondes, aircraft, and marine observations. Satellite tokenization uses a Transformer-VAE style encoder with masked attention; the pretraining objective is

$\hat{R}_{x,y}=[l,r,t,b]$1

with $\hat{R}_{x,y}=[l,r,t,b]$2. In-situ observations are mapped into a bird’s-eye-view anchor grid using point-transformer attention and altitude-conditioned normalization, then decoded back to arbitrary query points by kernel-weighted aggregation. Fusion across modalities is performed by a Multimodal Masked Autoencoder (MMAE) over a 12-hour sliding window, and temporal prediction is performed by a Transformer Forecast module trained with latent prediction loss

$\hat{R}_{x,y}=[l,r,t,b]$3

This architecture replaces several defining components of classical NWP. There is no explicit PDE solver, no gridded analysis state, and no variational or ensemble data-assimilation cycle. Instead, masked reconstruction across modalities implicitly assimilates information. In hindcast evaluation, the system reconstructs near-surface temperature with MAE $\hat{R}_{x,y}=[l,r,t,b]$4 K and V-component wind speed with MAE $\hat{R}_{x,y}=[l,r,t,b]$5 m/s at station locations, matching or surpassing ERA5 there. Forecast comparisons against IFS show identical MAE at early lead times for the examined surface variables, slight USpeed advantage for Earth-o1 in the $\hat{R}_{x,y}=[l,r,t,b]$6–$\hat{R}_{x,y}=[l,r,t,b]$7 h range, and continued advantage at $\hat{R}_{x,y}=[l,r,t,b]$8–$\hat{R}_{x,y}=[l,r,t,b]$9 h with USpeed MAE about $\hat{C}_{x,y}$0 m/s and dew point temperature MAE about $\hat{C}_{x,y}$1 K. For extreme precipitation at the 90th percentile threshold, Earth-o1 has a lower miss rate than IFS for the first 10 hours and then matches IFS thereafter. The fusion module also exhibits explicit scaling laws,

$\hat{C}_{x,y}$2

indicating systematic improvement with larger datasets and larger models. The current system is deterministic, atmosphere-centric, and does not yet include explicit ocean, land, hydrology, or biogeochemistry dynamics.

5. Unified 3D representation and scene generation

In generative 3D vision, OneWorld is a framework for single-image 3D scene generation that performs diffusion directly in a coherent 3D representation space rather than in 2D image or video VAE latents. The core module is the 3D Unified Representation Autoencoder (3D-URAE), built on the pretrained T$\hat{C}_{x,y}$3 3D foundation model. T$\hat{C}_{x,y}$4 begins with DINOv2 patch tokens

$\hat{C}_{x,y}$5

then uses image tokens and camera parameters $\hat{C}_{x,y}$6 to produce geometry tokens

$\hat{C}_{x,y}$7

and decodes them into 3D Gaussian Splatting parameters and depth maps,

$\hat{C}_{x,y}$8

OneWorld extends this geometry-centric pipeline by adding an appearance encoder $\hat{C}_{x,y}$9 and semantic distillation from DINOv2 into the unified 3D token space (Gao et al., 17 Mar 2026).

Appearance injection concatenates semantic and appearance features before geometry encoding,

$$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$0

and is empirically critical: removing appearance injection reduces reconstruction from PSNR $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$1 to $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$2, SSIM $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$3 to $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$4, and worsens LPIPS from $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$5 to $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$6. Semantic structure is imposed with two marginal alignment terms, a cosine loss and a distance-matrix similarity loss,

$$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$7

which are combined with rendering supervision,

$$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$8

The rendering loss itself combines pixelwise reconstruction and LPIPS on novel views.

Diffusion is then performed on flattened target-view tokens $$C^*_{x,y}= \sqrt{\frac{\min(l^*,r^*)}{\max(l^*,r^*)} \times \frac{\min(t^*,b^*)}{\max(t^*,b^*)}}.$$9. The denoiser is a DiT initialized from Wan-2.1-T2I-1.3B, conditioned on clean conditioning-view tokens, source and target camera parameters, and optional text embeddings. Although the model predicts clean tokens $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$0, training uses a velocity objective,

$$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$1

To preserve structural alignment across viewpoints, the method introduces token-level Cross-View-Correspondence (CVC) loss. For each predicted target token $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$2, cosine similarity is computed against all conditioning tokens; the nearest-neighbor index $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$3 is retained only when similarity exceeds threshold $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$4, and a cross-entropy loss is applied over the induced softmax distribution $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$5. The total diffusion loss is

$$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$6

with $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$7. A further Manifold-Drift Forcing (MDF) stage mixes ground-truth and diffusion-sampled latents,

$$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$8

and trains only the decoder so that off-manifold latent drift at inference does not destabilize 3DGS generation.

Quantitatively, the framework improves both 1-view novel-view synthesis and WorldScore-style consistency benchmarks. On RealEstate10K, OneWorld reaches PSNR $$\hat{I}_{x,y}=\mathrm{IoU}([l,r,t,b],[l^*,r^*,t^*,b^*]).$$9, SSIM $0.3$0, LPIPS $0.3$1, and Image Quality $0.3$2, outperforming FlashWorld and Gen3R. On DL3DV-10K, it reaches PSNR $0.3$3, SSIM $0.3$4, LPIPS $0.3$5, and Image Quality $0.3$6. In WorldScore-style evaluation, it attains 3D Consistency $0.3$7 and Photometric Consistency $0.3$8 on WorldScore-Indoor, and 3D Consistency $0.3$9 and Photometric Consistency $C^N$00 on DL3DV-style outdoor evaluation. Ablations show that removing CVC drops RealEstate10K performance from PSNR $C^N$01 to $C^N$02, and removing MDF reduces it to $C^N$03, confirming that explicit cross-view regularization and manifold shaping are not incidental additions but integral parts of the method.

6. Cross-domain synthesis, misconceptions, and limitations

The four usages of OneWorld differ sharply in domain, but they converge on a small set of architectural principles. First, each replaces a conventional decomposition with a more global substrate: FCOS-based OWP replaces proposal-conditioned two-stage pipelines with dense objectness and decoupled heads; OSWM replaces per-environment retraining with a single in-context transformer; Earth-o1 replaces grid-and-assimilation workflows with LatentOS plus MMAE and a learned latent forecast operator; OneWorld 3D replaces 2D latent diffusion with a unified 3D token manifold. Second, each uses this substrate to mediate multiple downstream tasks: localization plus classification, prediction plus policy learning, reconstruction plus forecasting plus inversion, or rendering plus novel-view synthesis plus text conditioning. This suggests that “OneWorld” consistently denotes shared-state modeling under partial supervision or partial observability rather than any single loss or model family (Konan et al., 2022, Ferreira et al., 2024, Gong et al., 7 May 2026, Gao et al., 17 Mar 2026).

A recurrent misconception would be to equate these systems with unrestricted universal world models. The detection work does not perform explicit unknown-class recognition or incremental learning; it evaluates recall of class-agnostic proposals on unseen classes. OSWM is not yet a general simulator: it adapts to GridWorld, CartPole-v0, and SimpleEnv, but transferring to MountainCar-v0 and Pendulum-v1 remains a challenge. Earth-o1 is not a complete Earth-system digital twin; it has no explicit dynamical ocean or land model, and uncertainty is not explicitly modeled. The 3D scene generator is not a dynamic 4D world model; it focuses on static scenes and single-image-conditioned scene generation.

The limitations are correspondingly domain-specific. In detection, Unknown Object Masking depends on the threshold $C^N$04 and the training stage at which masking begins, and its benefit is dataset-specific. In OSWM, the synthetic prior and context policy can fail to cover the state space needed for complex control, and reward smoothing can introduce incorrect incentives. In Earth-o1, physical consistency is learned rather than guaranteed, and sparse or poorly observed regimes remain difficult. In 3D generation, rare scene types, extreme viewpoints, and low-resolution training can still produce blurry textures or artifacts, while the computational budget remains substantial. A plausible implication is that future OneWorld systems will be judged less by the existence of a single latent space than by whether that space remains stable, informative, and operationally useful when scaled across heterogeneous modalities, longer horizons, and richer downstream control or inversion tasks.