WRBench/wrbench-results

WRBench Evaluation Results

Model-level diagnostic profile for Current World Models Lack a Persistent State Core — the aggregate table behind the WRBench leaderboard. This release covers 23 models on the 9,600-video public benchmark surface, each scored on six evaluation dimensions (D1-D6) under controlled viewpoint changes.

🏆 Explore it interactively in the WRBench Leaderboard.

Load

from datasets import load_dataset

scores = load_dataset("WRBench/wrbench-results", "model_scores", split="default")
df = scores.to_pandas().sort_values("Avg", ascending=False)

Metric definitions

All scores are in [0, 1]; higher is better.

Column	Paper dimension	What it measures
Cam Prec. (D1)	requested-camera precision	Strict requested-camera precision on certified local requested-control rows with explicit target trajectories. A separate diagnostic from Cam Align.; not applicable to prompt-camera / API rows.
Cam Align. (D1)	prompt-camera alignment	Common-yaw / static prompt-camera alignment diagnostic (CamAlign v2-lite). Not strict trajectory precision.
Integ. (D2)	visual integrity	Full-FoV DINOv2 visual-evidence / readability proxy.
Vis. spatial (D3)	visible spatial consistency	Spatial consistency evaluated while the relevant spatial evidence is observable.
Vis. state (D4)	visible state consistency	Action / state consistency evaluated while the relevant evidence is observable.
Reobs. Support	re-observation support	Fraction of outputs with judgeable hidden-and-returned evidence (the E14 applicability gate). Diagnostic coverage only: it establishes applicability/judgeability, not correctness.
Ret. spatial (D5)	returned spatial consistency	Spatial consistency conditional on judgeable re-observation evidence.
Ret. state (D6)	returned state consistency	State consistency conditional on judgeable re-observation evidence.
Avg	(aggregate)	Mean of Cam Align., Integ., Vis. spatial, Vis. state, Ret. spatial, Ret. state (paper main-table average).

Aggregation. Returned spatial/state metrics (D5/D6) are conditionally aggregated over the judgeable re-observation subset (the paper's shared-E14 pool of 2,073 rows); elsewhere they are null (NA). Outputs without identifiable returned evidence provide insufficient support for returned-state evaluation rather than counted failures. WRBench reports a diagnostic profile, not a single scalar ranking.

Viewpoint condition type. viewpoint_condition is the paper's viewpoint condition type (Table 2) — the form of input that supplies information about the requested viewpoint change:

• Source-video: a reference stream supplies appearance, layout, and partial event evidence.
• Geometry-cache: a point cloud / 3D / 4D record makes camera-target access computable.
• Model-inferred: no external view-state reference; the new view is synthesized under local camera/action/state controls.
• Prompt-only: natural-language camera intent only (uses CamAlign, not strict CamPrec).

Model input. model_input is the benchmark input interface delivered to each model (from the WRBenchLib inference contracts):

• TV2V (text + video): a source/reference video is the conditioning input (ReCamMaster, HyDRA, InSpatio World 14B, Gen3C, Spatia).
• TI2V (text + image): a text prompt plus a first-frame image is the conditioning input (all other models, including the prompt-only I2V APIs).

Schema

model_scores has one row per evaluated model:

Field	Type	Description
model_id	string	Internal model slug (matches wrbench-videos/videos/<model_id>/)
display_name	string	Public model name
viewpoint_condition	string	Paper viewpoint condition type: Source-video, Geometry-cache, Model-inferred, Prompt-only
model_input	string	Benchmark input interface: TV2V (text+video) or TI2V (text+image)
paper_group	string	Deprecated legacy grouping (V2V/Camera/Interactive/API prompt-camera); kept for backward compatibility — use viewpoint_condition
source_group	string	Provenance bucket (local_deployed, api_prompt_camera)
n_outputs	int	Evaluated outputs for this model
D1_CamPrec, D1_CamPrec_n	float, int	Requested-camera precision and its support count
D1_CamAlign_common_yaw, _n	float, int	Prompt-camera alignment (CamAlign v2-lite) and support
D1_CamAlign_static_hold, _n	float, int	Static-hold CamAlign component and support
D2-D6, D2_n-D6_n	float, int	Dimension scores and per-dimension support counts
gate_applicable_rate	float	Re-observation support rate (D5/D6 applicability gate)
low_n_flag	bool	True when a reported dimension has thin support
metric_notes	string	Per-row provenance / applicability caveats
Avg	float	Mean of CamAlign, D2, D3, D4, D5, D6 (paper main-table average)

NA indicates a dimension that is not applicable to a model (e.g. D1_CamPrec for API prompt-camera rows) rather than a zero score.

Provenance

• Benchmark date: 2026-06-08
• Commit: 73b804130
• CamAlign: d1_camalign_v2_lite (2026-06-10 common-yaw / static-hold intent)
• D5/D6: conditional over 2,073 shared-E14 re-observation-supported rows
• Raw exports kept under original/ (wrbench_23model_results.csv / .json)

Links

• 🏆 Leaderboard: https://huggingface.co/spaces/WRBench/wrbench-leaderboard
• 🎞️ Videos + per-video scores: https://huggingface.co/datasets/WRBench/wrbench-videos
• 🖼️ Natural-25 prompts: https://huggingface.co/datasets/WRBench/wrbench-natural25
• 🧑‍⚖️ Human annotations: https://huggingface.co/datasets/WRBench/wrbench-human-annotations
• 🧭 Release collection: https://huggingface.co/collections/WRBench/wrbench-current-world-models-lack-a-persistent-state-core-6a365c717251293c9fc2cc26
• 💻 GitHub: https://github.com/JinPLu/WRBench
• 🌐 Project page: https://jinplu.github.io/WRBench/
• 📄 Paper: https://arxiv.org/abs/2606.20545
• 🤗 HF paper page: https://huggingface.co/papers/2606.20545

Citation

@article{wrbench2026,
  title   = {Current World Models Lack a Persistent State Core},
  author  = {WRBench Team},
  journal = {arXiv preprint arXiv:2606.20545},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.20545}
}