WM ArenaAchievement timeline
Current seed summary renders from `time_alignment` and `metric_status`.
1Start: day
2End: day
3Daylight observed
4Night observed
5Status: ranking-adopted
Crafter Pilot
Pilot metrics, scripted expert, and rollout scaffolding
Crafter is in pilot mode. Δ-IRIS is the first planned evaluated model, while achievement extraction, spatial persistence, and seed-level rollout summaries are being staged behind this environment shell.
Δ-IRISfirst planned model
10 × 500seed/frame rollout target
1published Crafter evaluations
Inventory HUD
Current HUD summary renders from published Crafter pilot metrics.
achievement_f11.00
spatial_persistence1.00
metric_statusranking-adopted
ssim_t5000.37
Seed distribution
Published seed rows now render from `delta_iris.games[].crafter_metrics`.
S01
S10
S02
S03
S04
S05
S06
S07
S08
S09
achievement_f1
Pilot gate pendingUnlock event precision/recall across the 22 Crafter achievements.
spatial_persistence
Pilot gate pendingTracks whether revisited semantic tiles remain consistent over time.
ssim_t500
ExploratoryLong-horizon descriptive similarity at 500 frames.
Current manifest ships 10 published Crafter rollout seed row(s). This section can evolve without changing routes again. Focus model: Δ-IRIS.
Visual Comparison
Seed 01
SSIM 0.3748Δ-IRISNEW
Ground Truth
Δ-IRIS Prediction
0.3748
148.2
0.3266
500
Frames
View details →
Featured Games
Seed 10
0.3357JEDi—
Real
Predicted
Seed 02
0.5471JEDi—
Real
Predicted
Seed 03
0.4077JEDi—
Real
Predicted
Seed 04
0.3333JEDi—
Real
Predicted
Seed 05
0.2572JEDi—
Real
Predicted
Seed 06
0.416JEDi—
Real
Predicted
Seed 07
0.4035JEDi—
Real
Predicted
Seed 08
0.368JEDi—
Real
Predicted
Seed 09
0.3432JEDi—
Real
Predicted
Model Leaderboard
#
Model
Architecture
HNS
JEDi
Venue
Status
-
TWISTERBurchi, Timofte
Transformer + CPC
-
—
ICLR 2025
Pending
-
DIAMONDAlonso et al.
Diffusion (EDM)
-
—
NeurIPS 2024 Spotlight
Pending
-
EDELINELee, Lin, Sun, Lee
Mamba SSM + Diffusion
-
—
NeurIPS 2025
Pending
-
EfficientZero V2Ye et al.
MuZero + MCTS
-
—
ICML 2024 Spotlight
Pending
-
EfficientZeroYe et al.
MuZero + Self-Supervised
-
—
NeurIPS 2021
Pending
-
M³Cohen et al.
Modular Transformer
-
—
2025
Pending
-
HarmonyDreamMa et al.
RSSM + Harmony Loss
-
—
ICML 2024
Pending
-
OC-STORMZhang et al.
Object-Centric Transformer
-
—
2024
Pending
-
REMCohen et al.
RetNet + Parallel Observation
-
—
ICML 2024
Pending
-
EMERALDBurchi, Timofte
MaskGIT + Spatial Latent
-
—
ICML 2025
Pending
-
GIT-STORMMicheli et al.
Transformer + MaskGIT Prior
-
—
ICLR 2025
Pending
-
DreamerV3Hafner et al.
RSSM + Symlog
-
—
JMLR 2024
Pending
-
STORMZhang et al.
Stochastic Transformer
-
—
NeurIPS 2023
Pending
-
Δ-IRISMicheli, Alonso, Fleuret
Transformer + Delta Tokens
-
—
ICML 2024
Pending
-
DARTAgarwal, Andrews, Kahou
Fully Discrete Tokens
-
—
ICML 2024
Pending
-
IRISMicheli, Alonso, Fleuret
Transformer + dVAE
-
—
ICLR 2023
Pending
-
TWMChen et al.
Transformer + VQ-VAE
-
—
ICLR 2023
Pending
-
SimPLeKaiser et al.
VAE + Video Prediction
-
—
ICLR 2020
Pending
-
DramaWang et al.
Mamba-2 SSM
-
—
ICLR 2025
Pending
-
DreamerV2Hafner et al.
RSSM + Discrete Latent
-
—
ICLR 2021
Pending
-
BBFSchwarzer et al.
Scaled ResNet DQN
-
—
ICML 2023
Pending
-
SPRSchwarzer et al.
Rainbow DQN + Self-Prediction
-
—
ICLR 2021
Pending