Literature
Display
← All papers

JEPA-Anything: Learning Predictive Models across Different Worlds

Taoyong Cui · Zhongyao Wang · Xinyue Xu · Weiyang Liu · Zhaochen Yu · Yuying Zhang · Qiang Gao · Mengyue Yang · Wanli Ouyang · Pheng Ann Heng · Yingcheng Wu · Zhenfei Yin · Ling Yang

· arXiv (preprint) ↗

World models7 min text · CondensedOriginal paper ↗
Contents

Overview

The figure summarizes the shared predictive interface and its evaluation modes.

Refer to caption
Figure 1: Overview of JEPA-Anything. Upper section: Diverse domains share the same predictive problem: inferring unobserved, intervened, or future states from context. Standard JEPA mixes multiple predictive modes in a monolithic target embedding. In contrast, JEPA-Anything uses orthogonal predictive factorization (OPF) to allocate predictive capacity across complementary, orthogonal factors, each with a dedicated predictor, and synthesize a complete latent world state. The shared factorized core supports representation reuse, out-of-distribution and intervention prediction, long-horizon dynamics, and factor-level scientific analysis. Lower section: Three evaluation groups illustrate these capabilities. Group I (terminal readout) examines controlled visual binding and single-cell state representations. Group II (latent world dynamics) evaluates intervention-conditioned prediction and factor intervention in locomotion. Group III (scientific analysis) investigates wet-lab intervention nomination and the recovery of an orbital scaling law. Within each group, experiments A and B are arranged from left to right.

At a glance

JEPA-Anything is a common learning framework instantiated separately across seven domains, not one jointly trained universal model. It splits latent prediction into orthogonal factors and recombines them into a complete state. Matched experiments report improved prediction across ten dynamics tasks and four molecular systems. Planning improves in two of three locomotion environments. Scientific demonstrations examine preclinical biological interventions and orbital scaling.

1 Introduction

A single latent target can let easy or high-variance features dominate learning. OPF allocates prediction capacity to complementary subspaces while domain-specific adapters determine observations and context–target relationships. The learned state can support a terminal readout, repeated transitions, or factor-level analysis. Shared architecture does not imply shared cross-domain coordinates or causal meaning.

Related references: [12], [14], [3], [6], [5], [20], [8], [13].

2 JEPA-Anything

2.1 One Interface, One Predictive Core

Adapters turn each domain’s data into tokens and structural descriptors, then select a predictive context and target. An online encoder represents context; an EMA encoder supplies stopped-gradient targets. A valid task requires a meaningful statistical relationship between states of the same system.

Equations 1–2
x→𝒜δ(H,S)→𝒱δ(HC,SC,T,ST).x\xrightarrow{\mathcal{A}_{\delta}}(H,S)\xrightarrow{\mathcal{V}_{\delta}}(H_{C},S_{C},T,S_{T}). (1)
θ¯←m​θ¯+(1−m)​θ,0≤m<1,\bar{\theta}\leftarrow m\bar{\theta}+(1-m)\theta,\qquad 0\leq m<1, (2)
Table 1: What changes and what remains invariant across JEPA-Anything instantiations.
Component Domain-specific boundary Shared JEPA-Anything core
Raw representation Image, sequence, graph, set, record —
Tokenization and descriptors 𝒜δ\mathcal{A}_{\delta} Common token interface
Context–target semantics 𝒱δ\mathcal{V}_{\delta} Context predicts targets
Encoder architecture ViT, Transformer, GNN, MLP, …\ldots Online/EMA-target scheme
Predictive learning — Factorized predictors and state synthesis
Regularization — Orthogonality, factor activity, encoder variance
Training objective Original loss ℒbase(δ)\mathcal{L}_{\mathrm{base}}^{(\delta)} Additive OPF loss in Equation (10)
Downstream use Pooling, readout, probe, decoder, planner Online-encoder features or synthesized predicted states

2.2 Orthogonal Predictive Factorization

Learned projectors partition the target into K factors, each predicted by a dedicated branch. A pseudoinverse reconstructs the complete latent state for decoding or rollout. Within-factor and cross-factor orthogonality penalties discourage degenerate or redundant bases. Exact orthogonality preserves information and avoids amplification of coordinate error; arbitrary multi-head predictors lack that guarantee.

Related references: [36], [7].

Equations 3–8
z~t=sg(zt),zt(k)=Pk⊤z~t,k=1,…,K.\widetilde{z}_{t}=\operatorname{sg}(z_{t}),\qquad z_{t}^{(k)}=P_{k}^{\top}\widetilde{z}_{t},\qquad k=1,\ldots,K. (3)
z^t(k)=qk​(zc,st).\widehat{z}_{t}^{(k)}=q_{k}(z_{c},s_{t}). (4)
z^t=(P⊤)†​u^t∈ℝd,P=[P1,…,PK].\widehat{z}_{t}=\left(P^{\top}\right)^{\dagger}\widehat{u}_{t}\in\mathbb{R}^{d},\qquad P=[P_{1},\ldots,P_{K}]. (5)
u^t+1=[q1​(zt,ξt,st+1);…;qK​(zt,ξt,st+1)],z^t+1=(P⊤)†​u^t+1.\widehat{u}_{t+1}=\big[q_{1}(z_{t},\xi_{t},s_{t+1});\ldots;q_{K}(z_{t},\xi_{t},s_{t+1})\big],\qquad\widehat{z}_{t+1}=(P^{\top})^{\dagger}\widehat{u}_{t+1}.
ℒpred=1K​|T|​r​∑t∈T∑k=1K‖z^t(k)−zt(k)‖22,\mathcal{L}_{\mathrm{pred}}=\frac{1}{K|T|r}\sum_{t\in T}\sum_{k=1}^{K}\left\|\widehat{z}_{t}^{(k)}-z_{t}^{(k)}\right\|_{2}^{2}, (6)
ℒorth=∑k=1K‖Pk⊤​Pk−Ir‖F2+∑1≤i<j≤K‖Pi⊤Pj‖F2.\begin{split}\mathcal{L}_{\mathrm{orth}}={}&\sum_{k=1}^{K}\left\|P_{k}^{\top}P_{k}-I_{r}\right\|_{F}^{2}\\ &+\sum_{1\leq i<j\leq K}\left\|P_{i}^{\top}P_{j}\right\|_{F}^{2}.\end{split} (7)
‖P⊤​z‖22=∑k=1K‖Pk⊤​z‖22=‖z‖22,z=∑k=1KPk​Pk⊤​z.\|P^{\top}z\|_{2}^{2}=\sum_{k=1}^{K}\|P_{k}^{\top}z\|_{2}^{2}=\|z\|_{2}^{2},\qquad z=\sum_{k=1}^{K}P_{k}P_{k}^{\top}z. (8)
z^=P⁡(u+e),‖z^−z‖2=‖e‖2,κ2​(P)=1.\widehat{z}=P(u+e),\qquad\|\widehat{z}-z\|_{2}=\|e\|_{2},\qquad\kappa_{2}(P)=1.

2.3 Maintaining Factor and Encoder Activity

A factor-variance floor keeps projected coordinates active, while a separate variance term directly prevents the online encoder from collapsing. Factor prediction and regularization are added to each domain implementation’s original objective.

Equations 9–10
ℒfac\displaystyle\mathcal{L}_{\mathrm{fac}} =1K​r​∑k=1K∑j=1rmax⁡(0,γfac−σk,jfac),\displaystyle=\frac{1}{Kr}\sum_{k=1}^{K}\sum_{j=1}^{r}\max(0,\gamma_{\rm fac}-\sigma^{\rm fac}_{k,j}), (9)
ℒenc\displaystyle\mathcal{L}_{\mathrm{enc}} =1d​∑j=1dmax⁡(0,γenc−σjenc).\displaystyle=\frac{1}{d}\sum_{j=1}^{d}\max(0,\gamma_{\rm enc}-\sigma^{\rm enc}_{j}).
ℒOPF\displaystyle\mathcal{L}_{\mathrm{OPF}} =ℒpred+λorth​ℒorth+λfac​ℒfac+λenc​ℒenc,\displaystyle=\mathcal{L}_{\mathrm{pred}}+\lambda_{\mathrm{orth}}\mathcal{L}_{\mathrm{orth}}+\lambda_{\mathrm{fac}}\mathcal{L}_{\mathrm{fac}}+\lambda_{\mathrm{enc}}\mathcal{L}_{\mathrm{enc}}, (10)
ℒtrain(δ)\displaystyle\mathcal{L}_{\mathrm{train}}^{(\delta)} =ℒbase(δ)+ℒOPF.\displaystyle=\mathcal{L}_{\mathrm{base}}^{(\delta)}+\mathcal{L}_{\mathrm{OPF}}.

2.4 Shared Predictive Pretraining and Task-Specific Readout

All domains follow the same sequence: adapt observations, sample views, encode context and targets, predict factors, update online parameters, and update the target encoder by EMA.

Representation mode.

For ordinary representation tasks, retain the online encoder and attach a task-specific readout. Target encoders and factor predictors are no longer needed at inference.

Equation 11
hδ,τ​(x)=Rδ,τ​(fθδ​(𝒜δ​(x))),h_{\delta,\tau}(x)=R_{\delta,\tau}\!\left(f_{\theta}^{\delta}(\mathcal{A}_{\delta}(x))\right), (11)

Operational world-model mode.

For forecasting and planning, retain the factor predictors and synthesis map so predicted states can be reused recursively. The projectors also expose coordinates for intervention and spectral analyses.

Equation
uk(x)=Pk⊤zδ(x),k=1,…,K.u_{k}(x)=P_{k}^{\top}z_{\delta}(x),\qquad k=1,\ldots,K.

3 Experiments

Evaluation distinguishes terminal representations, recursive dynamics and planning, and scientific analysis of learned factors.

3.1 Common Evaluation Logic

Direct JEPA comparisons hold adapters, encoders, data splits, optimization budgets, and readouts fixed. Domain-specific metrics are reported separately rather than averaged across incompatible units. Control uses five training seeds and matched resets.

Table 2: Overview of JEPA-Anything instantiations and evaluation protocols. In Groups I and II, each implementation combines its base loss with ℒOPF\mathcal{L}_{\mathrm{OPF}} as specified in Equation (10). Numerical OPF entries use the compact form d;K×rd;\,K\!\times\!r, where K​r=dKr=d; symbolic entries denote domain-configured widths. Group III reports the factor interface used for analysis and external validation.
Group/domain and data Context →\rightarrow target or analysis Backbone or interface OPF setup / interface Seeds / splits / metrics
I / vision; controlled MuJoCo scenes Visible patches →\rightarrow masked patch states DINOv3 ViT-S SigLIP2 Base 384;×96384;\,4\!\times\!96 768;×192768;\,4\!\times\!192 10 seeds; image-disjoint and leave-one-cell-out; INJ, Coll., Rec.
I / single cell; kidney, PBMC-10K, Adamson, Norman Representation: masked expression →\rightarrow complete cell state; Norman: target zresponse:=zobserved−zcontrolz_{\mathrm{response}}:=z_{\mathrm{observed}}-z_{\mathrm{control}} scGPT Norman: 512;×128512;\,4\!\times\!128 5 seeds; kidney pretraining; PBMC fine-tuned/zero-shot and cross-dataset perturbation transfer; AvgBIO, Pearson
I / clinical; longitudinal multimodal cohort Patient history →\rightarrow future patient state GPT-2 small; risk decoder 768;×192768;\,4\!\times\!192 5 seeds; fixed patient-level cohort split; more than 1,000 event risks; PRAUC
II / intervention; CITRIS Interventional Pong Observation + intervention →\rightarrow next state Matched encoder and predictor 160;×32160;\,5\!\times\!32 5 paired seeds; single- and combined-intervention one-step prediction; six-step free rollout; four-channel MSE
II / dynamics; CausalWorld, DMC, PDEBench, WeatherBench2 State, frames, action, or forcing →\rightarrow next state or field Matched task-specific encoder and transition 128;×32128;\,4\!\times\!32 5 seeds; held-out/OOD and 6-step rollout; MSE and return
II / PDE audit; APEBench Burgers and KS Current field →\rightarrow next field MLP; first-layer weight ×64256\!\times\!64 128;×32128;\,4\!\times\!32 5 paired seeds; held-out late-state prediction; Burgers 6-step rollout; MSE
II / long rollout; 1-D viscous Burgers, 64 points Current field →\rightarrow autoregressive future field MLP; input xdim=1024x_{\rm dim}=1024; hidden width 256 128;×32128;\,4\!\times\!32 3 seeds; 2,000 training steps; ID/high-frequency OOD at H20/H50; MSE
II / locomotion; Hopper, Walker2d, HalfCheetah State + action →\rightarrow future control state Latent dynamics; CEM planner 32;×832;\,4\!\times\!8 5 seeds; matched resets and 20-step factor interventions; return, R2R^{2}, MSE
II / molecules; water, quartz, paracetamol, benzene Atomic state →\rightarrow future molecular state TrajCast-style O⁡(3)O(3) equivariant; l=1l=1 sector 64;×1664;\,4\!\times\!16 (channels) 5 seeds; 100-step free rollout; MAE, RMSD
III / biology; Huh7–PBMC, organoids, tumor fragments, mice Factor coordinates →\rightarrow candidate intervention nomination OPF factor-analysis interface Retained PkP_{k} 3 organoids and 3 fragments plus external mouse validation; killing and activation
III / orbital; simulated position–velocity trajectories Factor coordinates →\rightarrow latent spectral modes Trajectory encoding; spectral analysis Retained PkP_{k} One analyzed run with analytic-law validation; slope and R2R^{2}
Refer to caption
Figure 2: Scenario atlas for JEPA-Anything. The layout follows the paper’s three evaluation groups: terminal readout, latent world dynamics, and scientific analysis. Domain-specific contexts and outputs are connected through the same factorized predictive-state interface; Group III retains the learned factor coordinates for external biological validation and physical-law analysis.

3.2 Group I: Terminal Readout of Predictive World States

Vision and cell tasks read the trained encoder; disease forecasting reads a synthesized future state. These tests terminate at a readout rather than repeatedly feeding predictions back.

3.2.1 Controlled Visual Binding

Question.

The visual task asks whether frozen features retain both the location and type of an image change, including combinations excluded from readout training.

Predictive pretraining.

DINOv3 and SigLIP2 checkpoints are pretrained with matched masked-patch tasks. Four-factor OPF replaces the monolithic JEPA target; the resulting encoders are then frozen.

Related references: [27], [33].

Readout and evaluation.

Image-disjoint, held-out-cell evaluation measures binding accuracy, collapse, and recovered grid structure over ten seeds. OPF improves both backbones under the same learned-grid readout.

Related references: [32], [19].

Table 3: Controlled visual binding on MuJoCo, separating encoder pretraining from downstream grid supervision. Known grid supplies the support–operation taxonomy; learned grid estimates it from flat cell labels. Standard JEPA and JEPA-Anything use the same learned-grid readout.
Backbone Encoder pretraining Grid supervision INJ ↑\uparrow Coll. ↓\downarrow Rec. ↑\uparrow
DINOv3 Frozen checkpoint Known grid .452 .667 —
DINOv3 Frozen checkpoint Learned grid .569 .433 .643
DINOv3 Standard JEPA Learned grid .572 .426 .645
DINOv3 JEPA-Anything Learned grid .581 .417 .659
SigLIP2 Frozen checkpoint Known grid .476 .656 —
SigLIP2 Frozen checkpoint Learned grid .484 .511 .676
SigLIP2 Standard JEPA Learned grid .483 .514 .679
SigLIP2 JEPA-Anything Learned grid .490 .503 .688

3.2.2 Single-Cell State Representation and Perturbation Prediction

Question.

Latent prediction aims to retain cellular state while reducing sensitivity to sparse expression counts and technical dropout.

Protocol.

The cell experiments use a Cell-JEPA-style masked scGPT encoder. Perturbation prediction factorizes the response relative to a control cell into four 128-dimensional factors before decoding expression changes.

Related references: [10], [9].

Evaluation.

Pretraining uses about 800,000 kidney cells. Evaluation covers PBMC clustering and Adamson/Norman perturbation prediction over five seeds; the factorized model leads the compared metrics in the reported table.

Related references: [1], [24].

Table 4: Single-cell representation and perturbation evaluation. Perturbation columns report Pearson correlation on absolute post-perturbation expression.
Model JEPA Orthogonal factors PBMC finetuned AvgBIO ↑\uparrow PBMC zero-shot AvgBIO ↑\uparrow Norman Pearson ↑\uparrow Adamson Pearson ↑\uparrow
scGPT No No 0.7531 0.5288 0.631 0.905
Cell-JEPA Yes No 0.7830 0.7194 0.787 0.937
JEPA-Anything Yes Yes 0.8301 0.7752 0.814 0.942

3.2.3 Longitudinal Disease-State Forecasting

Question.

Clinical forecasting tests whether a latent future state predicts events from sparse, irregular patient histories.

Protocol.

A GPT-2-small encoder represents past history and an EMA encoder represents future states. Four factors reconstruct a future representation used by a shared decoder for more than 1,000 event risks.

Related references: [25].

Evaluation.

Mean PRAUC across the event vocabulary improves over a matched monolithic JEPA on a fixed patient-level split and five training seeds. This is retrospective predictive evaluation.

Refer to caption
Figure 3: Broad-spectrum prediction of more than 1,000 future clinical events. Methods are ranked by mean PRAUC using the same cohort split and event vocabulary; higher is better.

3.3 Group II: Latent World Dynamics and Rollout Stability

Recursive prediction tests whether errors remain controlled when synthesized latent states drive later forecasts or planning decisions.

3.3.1 Intervention-Conditioned Compositional Prediction

Question.

Interventional Pong supplies a pre-intervention state and labels of the changed factors. Evaluation covers seen single interventions, unseen combinations, and six-step free rollout.

Related references: [21].

Protocol and evaluation.

Against the matched dense JEPA, mean MSE falls from 0.009541 to 0.006218 for single interventions, 0.009441 to 0.008223 for unseen combinations, and 0.009478 to 0.008665 for six-step rollout: reductions of 34.83%, 12.90%, and 8.58%.

Refer to caption
Figure 4: Intervention-conditioned prediction and free rollout on CITRIS Interventional Pong. Lower four-channel MSE is better. Points show means over five paired seeds; annotations give the reduction computed from the reported means.

Orthogonality mechanism audit.

A capacity-matched multi-head audit measures subspace overlap, singular values, conditioning, synthesis error, and factor activity. Its exact-coordinate reconstruction test isolates geometry from prediction error; it is distinct from the forecasting benchmark.

Table 5: Mechanism audit of the factor interface on CITRIS Interventional Pong. Entries are mean ±\pm sample standard deviation over five paired seeds.
Metric Unconstrained multi-head Orthogonal factorization
Cross-factor subspace overlap ↓\downarrow 0.4550±0.04200.4550\pm 0.0420 (5.18±0.34)×10−16(5.18\pm 0.34)\!\times\!10^{-16}
Minimum singular value σmin​(P)\sigma_{\min}(P) ↑\uparrow 0.00513±0.002330.00513\pm 0.00233 0.999989±0.0000030.999989\pm 0.000003
Condition number κ2​(P)\kappa_{2}(P) ↓\downarrow 438.52±170.00438.52\pm 170.00 1.00005±0.0000041.00005\pm 0.000004
Transpose synthesis NMSE (exact factors) ↓\downarrow 0.7886±0.02990.7886\pm 0.0299 (2.98±0.11)×10−14(2.98\pm 0.11)\!\times\!10^{-14}

3.3.2 Out-of-Distribution and Long-Horizon Dynamics

Question.

The ten-task dynamics suite tests unseen initial states, episodes, or periods and then six-step recursive prediction across visual, control, physical-field, and weather systems.

Related references: [2], [30], [29], [26].

Protocol.

Both variants share data, backbones, budgets, splits, and five seeds within each task. Rollouts use the true action or target sequence. Weather normalization uses training-period statistics only.

Refer to caption
Figure 5: Results from the ten-task matched dynamics benchmark. Panel (a) shows the relative MSE reduction of JEPA-Anything versus standard JEPA, computed from the reported five-seed means for the nine prediction tasks. Panel (b) shows CausalWorld closed-loop control return as mean ±\pm sample standard deviation over five seeds; higher return is better.

Rollout and factor diagnostics.

The original table reports first- and sixth-step errors for representative systems, exposing error accumulation.

Table 6: Six-step rollout diagnostics. Lower MSE is better. Each entry is averaged over five seeds; the full evaluation records every intermediate step.
Standard JEPA JEPA-Anything
Task Step 1 Step 6 Step 1 Step 6
CausalWorld state prediction 0.01306 0.01307 0.01037 0.01079
DMC pixel dynamics 0.008539 0.008594 0.008498 0.008526
PDEBench Burgers 0.001830 0.006369 0.001101 0.004014
PDEBench shallow water 0.007090 0.010510 0.003999 0.006522

Additional APEBench evaluation.

Under the separate APEBench protocol, Burgers errors decrease about 49.5% on held-out late states and 44.7% over six steps; KS late-state error decreases about 13.2%. Improvements occur in every tested seed.

Related references: [18].

Table 7: Prediction results in the additional APEBench evaluation. Lower MSE is better.
System Metric Standard JEPA JEPA-Anything
Burgers Held-out late MSE ↓\downarrow 0.16621±0.017400.16621\pm 0.01740 0.08389±0.017700.08389\pm 0.01770
Burgers Six-step rollout MSE ↓\downarrow 0.29186±0.006820.29186\pm 0.00682 0.16153±0.035000.16153\pm 0.03500
Kuramoto–Sivashinsky Held-out late MSE ↓\downarrow 1.1012±0.07631.1012\pm 0.0763 0.9558±0.03400.9558\pm 0.0340

Capacity-matched Burgers long-rollout evaluation.

A capacity-matched Burgers experiment extends evaluation to 50 steps and high-frequency OOD trajectories. Gains persist but narrow with horizon. Routing diagnostics across nine tasks show about 3.782 effective factors out of four.

Table 8: Capacity-matched Burgers long-rollout evaluation over three seeds. Lower error is better. Relative change is computed against standard JEPA.
Split Standard H20 / H50 JEPA-Anything H20 / H50 JEPA-Anything relative change
ID 1.1185 / 0.8181 1.0532 / 0.7923 −5.84%-5.84\% / −3.15%-3.15\%
High-frequency OOD 1.0649 / 0.8452 1.0116 / 0.8214 −5.01%-5.01\% / −2.82%-2.82\%

3.3.3 Continuous-Control Planning

Question.

A CEM planner scores candidate action sequences through the learned model, then executes the first action of the best sequence. Real-environment return measures whether improved predictions lead to better decisions.

Protocol.

Matched Hopper, Walker2d, and HalfCheetah experiments use five training seeds and matched environment resets and planner randomness. OPF improves mean return on Walker2d and HalfCheetah, while standard JEPA performs better on Hopper. Benefits are environment-dependent.

Related references: [32].

Refer to caption
Figure 6: Capacity-matched continuous-control planning. Squares show the mean paired CEM-return difference between JEPA-Anything and standard JEPA over five training seeds; error bars show 95% confidence intervals. Positive values favor JEPA-Anything.

3.3.4 Factor Activity and Functional Interventions in Locomotion

Question and protocol.

Factor probes, activity measurements, and interventions test whether each of four eight-dimensional state subspaces carries useful locomotion information.

Descriptive readout.

Individual factors show different linear-readout preferences for body height and joint angles. These are learned statistical preferences rather than fixed semantic labels.

Activity and intervention.

All factors remain active. Replacing any one by its training-set mean increases 20-step rollout error in all three environments; planning return also often declines. This supports functional use of each factor without establishing causal disentanglement.

Refer to caption
Figure 7: Factor-wise functional interventions. Each factor is replaced by its training-set mean throughout evaluation. Panel (a) gives the increase in 20-step rollout MSE. Panel (b) gives the HalfCheetah CEM-return change as mean ±\pm sample standard deviation over five training seeds.

3.3.5 Force-Free Molecular Forecasting

Question.

Force-free molecular forecasts recursively predict positions and velocities, making long-horizon error an essential complement to one-step accuracy.

Protocol.

Matched equivariant models are evaluated on water, quartz, paracetamol, and benzene. Factorization acts on feature multiplicity channels identically across vector components, preserving rotational equivariance.

Related references: [31].

Evaluation.

Across five seeds, JEPA-Anything has the lowest one-step displacement MAE and median final-position RMSD after 100 steps among the compared methods in all four systems.

Table 9: Force-free molecular forecasting with a TrajCast-style backbone. Entries are means over five trained seeds. For each seed, we compute one-step displacement MAE and median final-position RMSD after 100 autoregressive steps. Values are reported in Å; lower is better.
System Metric ↓\downarrow Scratch TrajCast-JEPA JEPA-Anything
Water MAE 0.00387 0.00452 0.00376
Water RMSD 3.331 2.536 2.459
Quartz MAE 0.01080 0.01043 0.01011
Quartz RMSD 2.089 1.912 1.877
Paracetamol MAE 0.00901 0.00777 0.00705
Paracetamol RMSD 3.155 1.868 1.776
Benzene MAE ×10−52.59\!\times\!10^{-5} ×10−52.10\!\times\!10^{-5} ×𝟏𝟎−𝟓\mathbf{2.05\!\times\!10^{-5}}
Benzene RMSD 0.0958 0.0701 0.0645

3.4 Group III: Scientific Analysis of Learned World States

The scientific demonstrations connect learned factor coordinates to biological experiments and a known physical scaling law.

3.4.1 Wet-Lab Evaluation of a Factor-Nominated Intervention

Factor analysis nominated IL-18 with NT5E/CD73 blockade. The authors report enhanced antitumor activity in co-cultures, patient-derived organoids and tumor fragments, and mice. These are preclinical findings, not evidence of clinical efficacy.

Refer to caption
Figure 8: Selected wet-lab results in patient-derived hepatocellular carcinoma organoids and tumor fragments. Across three organoid specimens and three tumor-fragment specimens, the IL-18 plus CD73 blockade combination showed the strongest tumor-cell killing among the tested conditions, together with increased immune-cell activation. Colors identify blank control, IL-18, CD73 inhibition, and the combined intervention.

3.4.2 Orbital Scaling as a Physical-Law Diagnostic

Spectral analysis of simulated orbital trajectories recovers the frequency–semimajor-axis exponent at −1.4991, close to the Keplerian value −1.5. This tests whether latent modes retain a known physical regularity.

Refer to caption
Figure 9: A physical-law diagnostic based on predictive latent modes. In a JEPA-Anything run, the learned spectral modes recover the Keplerian ridge f=(2π)−1a−3/2f=(2\pi)^{-1}a^{-3/2} with fitted slope −1.4991-1.4991 and R2=0.9999999R^{2}=0.9999999.

4 Related Work

Latent world models and planning.

Earlier latent world models primarily develop representations and transitions for visual prediction or control; this work applies a common factorized objective across broader domains.

Related references: [12], [14], [15], [37], [4].

Joint-embedding predictive learning.

Joint-embedding methods predict representations rather than raw observations. Their learned invariances still depend on the selected context–target relation.

Related references: [3], [6], [4], [10], [28].

Structured and factorized representations.

Object slots and causal factors require appropriate structure or supervision. OPF does not assume its predictive subspaces correspond to independent semantic causes.

Related references: [16], [23], [17], [21], [22].

Redundancy reduction and collapse prevention.

The method extends redundancy reduction with constraints on target-subspace bases, explicit activity terms, and a synthesis map for complete latent states.

Related references: [36], [7].

5 Summary

A shared factorized objective supports several domain-specific prediction pipelines. Results favor it across many forecasting tasks, with mixed planning outcomes and separate biological and physical demonstrations.

6 Future Work

Future scientific-agent applications require calibrated uncertainty, reliable experiment selection, and revision under new evidence. Orthogonality alone establishes neither causality nor aligned representations across domains.

Related references: [35], [11], [34].

Appendix A Approximate Pseudoinverse-Synthesis Stability

When the learned basis is approximately orthogonal, its singular values bound the amplification of factor-prediction error. Pseudoinverse synthesis exactly reconstructs exact coordinates when the basis is invertible; the displayed proof states the assumptions.

Equations 12–13
1−ϵ≤σi​(P)≤1+ϵ,κ2​(P)≤1+ϵ1−ϵ.\sqrt{1-\epsilon}\leq\sigma_{i}(P)\leq\sqrt{1+\epsilon},\qquad\kappa_{2}(P)\leq\sqrt{\frac{1+\epsilon}{1-\epsilon}}. (12)
z^=(P⊤)†​u^=z+(P⊤)−1​e,‖z^−z‖2≤11−ϵ​‖e‖2.\widehat{z}=(P^{\top})^{\dagger}\widehat{u}=z+(P^{\top})^{-1}e,\qquad\|\widehat{z}-z\|_{2}\leq\frac{1}{\sqrt{1-\epsilon}}\|e\|_{2}. (13)

Appendix B Supplementary Reproducibility Information

B.1 Implementation Details

The physical-field models use four factors in a 128-dimensional latent state; the Norman cell-response model uses four factors in 512 dimensions.

B.2 OPF Coefficients

The preserved table records factor dimensions and regularization coefficients by experiment.

Table 10: OPF dimensions and coefficients of the three additive regularizers in Equation (10).
Experiment dd KK rr λorth\lambda_{\mathrm{orth}} λfac\lambda_{\mathrm{fac}} λenc\lambda_{\mathrm{enc}}
CITRIS Interventional Pong 160 5 32 .10 .05 .02
CausalWorld state/control 128 4 32 .10 .05 .02
DMC pixel dynamics 128 4 32 .10 .05 .02
DMC-VB held-out episodes 128 4 32 .10 .05 .02
PDEBench Burgers 128 4 32 .10 .05 .02
PDEBench reaction–diffusion 128 4 32 .10 .05 .02
PDEBench shallow water 128 4 32 .10 .05 .02
Synthetic PDE 128 4 32 .10 .05 .02
Synthetic pixel dynamics 128 4 32 .10 .05 .02
WeatherBench2 128 4 32 .10 .05 .02
APEBench Burgers / Kuramoto–Sivashinsky 128 4 32 .10 .05 .02
Hopper-v5, 20-step 32 4 8 .02 .10 .02
Walker2d-v5, 20-step 32 4 8 .02 .10 .02
HalfCheetah-v5, 20-step 32 4 8 .02 .10 .02

B.3 AI Use Statement

The authors disclose AI assistance for editing, manuscript preparation, checks, and experimental feedback, with author review and responsibility for the reported claims.

B.4 Ethics Statement

The clinical study uses de-identified UK Biobank data under an approved research application; participant-level data are not redistributed.

B.5 Reproducibility Statement

The paper specifies objectives, adapters, matched controls, splits, seeds, and coefficients. Non-OPF settings are shared within each direct comparison.

Appendix C Author Organizations and Correspondence

Organizations

The collaboration includes PhAI Labs and universities in Hong Kong, China, the United Kingdom, and the United States.

Correspondence

The original paper lists the corresponding authors and contact addresses.

Concept demonstration

Open image ↗