Current World Models Lack a Persistent State Core

Jinpeng Lu; Dexu Zhu; Haoyuan Shi; Linghan Cai; Guo Tang; Yinda Chen; Jie Cao; Duyu Tang; Yi Zhang; Yong Dai; Xiaozhu Ju

Source-linked AI summary

Current World Models Lack a Persistent State Core

Jinpeng Lu, Dexu Zhu, Haoyuan Shi, Linghan Cai, Guo Tang, Yinda Chen, Jie Cao, Duyu Tang, Yi Zhang, Yong Dai, Xiaozhu Ju

arXiv:2606.20545v1Jun 18, 2026cs.CV

TL;DR

Existing world-model benchmarks do not test whether scenes continue evolving while unobserved. WRBench evaluates this through viewpoint interventions and finds that current systems do not reliably preserve event endpoints when targets are re-observed.

Problem
Existing benchmarks assess fidelity, motion, and camera controllability but do not test whether dynamic world states continue evolving independently of observation.
Method
WRBench uses viewpoint intervention and a hierarchical six-dimension diagnostic chain to evaluate camera execution, visible consistency, re-observation, and returned-state consistency.
Results
Across 23 models and 9,600 videos, systems did not reliably maintain event endpoints after targets were re-observed, separating re-observed-state consistency from visible fidelity and access.
Takeaways & Limitations
Persistent world models should prioritize endpoint binding and consistency of event-induced states under changes in observability.
Takeaways & Limitations
Endpoint-directed reward or policy training remains proposed rather than demonstrated within the evaluated work.

Abstract

from arXiv · show

World models are increasingly regarded as a decisive step toward artificial general intelligence, yet modeling the physical world demands more than rendering convincing frames on demand: it requires an internal world state that keeps evolving over time, decoupled from observation, so that objects endure and events run to their conclusions whether or not a camera is watching, much as the moon holds to its orbit when no one is looking. This requirement is a blind spot of existing benchmarks, which reward surface properties such as fidelity, motion, and camera controllability while never asking whether a generated world keeps evolving once it is unobserved. We introduce \textbf{WRBench}, the first systematic diagnostic benchmark that treats camera motion as an intervention on observability and resolves evaluation into a human-calibrated chain that asks whether the camera executes the requested interaction, whether the scene stays continuous and identifiable while in view, and whether a returning target remains consistent with the event that was set in motion. Across 9{,}600 videos from 23 models spanning four control paradigms, one finding proves stubborn: current systems maintain the observed world as a tracking shot, resuming a returning target in the state at which it was abandoned rather than advancing the event while it went unseen. Because this failure recurs across control paradigms, model families, and increments of scale, robust world-state evolution does not follow from cleaner imagery, tighter control, richer geometric priors, or sheer parameter count We therefore argue that the stability of the physical state kernel and the consistency of worldlines under viewpoint intervention should become first-class objectives of world-model design, so that a world model captures how the world will unfold rather than how the next frame appears.

1 Introduction

WRBench frames world modeling as persistent state evolution under viewpoint intervention, not merely realistic frame generation. It diagnoses whether events continue while unobserved and whether returning evidence preserves the event endpoint.

Introduction: World models must maintain a continuously evolving state independent of observation, a capability existing benchmarks do not directly test.Prior benchmarks assess fidelity, motion, and camera controllability without asking whether a dynamic world keeps evolving when unobserved.
Introduction: Viewpoint change tests persistence by altering observation while leaving the world unchanged: a cat jumping onto a bed should be on the bed when the camera returns.The returning target must reflect the event’s endpoint rather than merely reappear plausibly.
Introduction: WRBench combines 25 scene families with four-level event designs and treats camera control as an intervention on observability.Each case specifies an initial scene, event, and viewpoint intervention, then evaluates evidence about the evolving world state.
Introduction: Its six-dimension diagnostic chain progresses from camera execution and visible consistency to re-observation access and returned spatial-state consistency.Re-observed-consistency scores are computed only for cases passing the re-observation gate.
Introduction: 2,547 deduplicated human annotator verdicts calibrate WRBench’s automatic evaluators, while WRBenchLib records delivered viewpoint-control conditions across heterogeneous generators.The toolkit supports comparisons across source-video, geometry-cache, model-inferred, and prompt-only controllability paradigms.
Introduction: 9,600 videos from 23 generators show that re-observed-state consistency is distinct from visible fidelity and re-observation access, and is not bound by scale within the evaluated systems.The results identify a recurring preservation–access–re-observed-consistency gap and motivate what-memory and endpoint-persistence objectives.

2 Related Work and Positioning

WRBench extends decomposed video and world-generation evaluation by treating camera control as an evidence intervention on dynamic world-state consistency. It positions itself as a complementary, human-calibrated diagnostic benchmark for whether event-induced states remain supported across visibility, interruption, and re-observation.

Decomposed video evaluation: Prior decomposed-video suites separate video quality, motion, and text–video alignment into interpretable measurements with tailored evaluators.FETV and EvalCrafter organize these dimensions, while VBench adds human validation and reusable evaluation infrastructure.
World-generation and physical evaluation: World-generation benchmarks probe prompt composition, event order, metamorphic structure, physical commonsense, physical-law adherence, instruction following, and camera or layout specification.The cited suites include T2V-CompBench, TC-Bench, ChronoMagic-Bench, StoryEval, VideoPhy, VideoPhy-2, PhyGenBench, T2VPhysBench, WorldModelBench, and WorldScore.
WRBench positioning: WRBench treats camera control as an evidence intervention and asks whether the same event-induced state remains supported while visible, temporarily unobserved, and returned to a judgeable view.It inherits VBench’s human-calibrated discipline and WorldScore’s control-aware world-generation framing while targeting dynamic world-state attribution under viewpoint intervention.
State, memory, and out-of-sight dynamics: Related benchmarks examine continuity under interrupted observation, out-of-sight dynamics, event permanence, memory, interaction, and revisiting.STEVO-Bench, LiveWorld, MBench, MIND, WorldMark, WBench, and iWorld-Bench cover these neighboring concerns.
Heterogeneous viewpoint condition types: Recent video generators expose heterogeneous viewpoint interfaces, making cross-interface evaluation itself a challenge.Examples span explicit camera trajectories, novel-view or multi-view generation, source-video transformation, spatial memory, 4D geometric control, and out-of-sight memory.

3 WRBench Suite

WRBench evaluates dynamic world-state behavior as evidence of a specified event under viewpoint intervention, rather than as standalone visual quality. Its four-layer pipeline standardizes event-view inputs across heterogeneous generators and separates camera execution, visual integrity, visible consistency, re-observation support, and returned-state consistency.

Suite Design: WRBench comprises Natural-25, WRBenchLib, a six-dimensional evaluation suite, and human preference annotation that calibrates automatic evaluators axis by axis.Natural-25 defines the scene and event substrate; WRBenchLib delivers test conditions and records outputs and provenance.
Evaluation Unit: The event-view record—not a prompt or clip alone—fixes the scene, event, viewpoint intervention, visibility regime, and interface variant used for evaluation.Natural-25 crosses 25 scene families with a four-level event design factoring spatial displacement against state change.
Generation Interface: WRBenchLib translates each event-view record into the form each generator can consume and logs the delivered condition, generated video, and provenance for reproducibility.This avoids forcing heterogeneous models into one control interface and compares them on the evidence they actually received.
Diagnostic Dimensions: The six dimensions progress from requested or intended camera behavior and visual integrity to visible spatial/state consistency, re-observation support, and returned spatial/state consistency.The final consistency test is conditional on judgeable return evidence; missing support is recorded as insufficient evidence rather than success or failure.
Diagnostic Dimensions: The diagnostic profile keeps view failure, unreadable evidence, visible event error, missing re-observation support, and returned-state error as distinct outcomes.This preserves interpretable failure modes instead of averaging them away.

4 Experiment

Across 23 models, experiments show that world-model controllability improves access and visible quality without producing persistent hidden-state evolution. Re-observed-state correctness remains the central failure, especially for in-place transformations, indicating that endpoint persistence requires dedicated architectural and training objectives.

4.1.1 Findings: Re-observed-state correctness is distinct from visible fidelity and camera access, which correlate weakly with whether a returned object has the correct state.Visible spatial and state scores correlate at r=0.97, while the re-observed pair correlates at r=0.94; the two blocks track each other only moderately.
4.1.1 Findings: Object transformation is harder than relocation: in-place changes reduce re-observed spatial by −0.075 and re-observed state by −0.068, while relocation benefits pretrained motion priors.In-place changes also reduce visible position by −0.114 while leaving visible state essentially flat at −0.031.
4.1.1 Findings: At matched return rates of 40.5% versus 40.4%, in-place events still score lower on visible position and state, at 0.661/0.502 versus 0.819/0.629.The per-model interaction term on re-observed state is indistinguishable from zero, with mean −0.009 and Wilcoxon p = 0.45.
4.1.2 Paradigms: More source footage increases re-exposure access, but no paradigm resolves hidden-state consistency: Gen3C reaches 0.711 versus 0.559, InSpatio 0.720 versus 0.591, and Spatia 0.633 versus 0.512.Prompt-only access is low, with Hailuo at 6.3% and Kling at 3.3%, while Wan-Fun A14B reaches 17.6%.
4.1.3 Scaling: Scaling and version upgrades improve observable quality but leave conditional re-observed state in a fixed band, so standard video scaling does not deliver endpoint persistence.Wan 1.3B→14B lifts the Wan-Bench weighted score 0.689→0.724 and the VBench total 83.96%→86.22%.
4.1.4 Design implications: The experiments identify a missing state carrier: architectures need a what-memory that records hidden changes, followed by long-horizon persistence training and explicit camera-execution supervision.The proposed long-to-short recipe learns persistence on long horizons before adding camera-execution supervision.

5 Conclusion · Appendix · A Use of Large Language Models (LLMs)

WRBench evaluates whether video-generation models maintain evolving world state under camera-induced partial observation, and across 23 models and 9,600 outputs finds that models do not reliably preserve event endpoints after re-observation. The appendix documents LLM use and provides sections on reproducibility, limitations, metric checks, failure forensics, model reading, human validation, and benchmark extensions.

5 Conclusion: WRBench tests whether video-generation models maintain evolving world state under camera-induced partial observation.
5 Conclusion: 23 models and 9,600 generated outputs reveal a preservation–access–re-observed-consistency gap in maintaining event endpoints.Models may preserve visible evidence, execute plausible camera motion, or expose the target region again without reliably maintaining the event endpoint.
5 Conclusion: Endpoint binding is the shared bottleneck because event-induced object relations, contact endpoints, postures, containment, or collision endpoints must be written back into the returned scene.
Appendix: The appendix includes sections on LLM use, reproducibility and metric–dataset records, limitations and future work, and metric–dataset common-problem checks.
Appendix: Additional appendix sections cover visual-integrity implementation, frame-level failure forensics, model subtype and series guidance, and human validation notes.
Appendix: The appendix also documents preference-pair export and reward/policy outlook, including dense-control and future benchmark extensions.
A Use of Large Language Models (LLMs): Large language models assisted with manuscript organization, wording, LaTeX editing, and consistency checks.Authors retained responsibility for scientific claims, definitions, numerical results, tables, figures, and citations.
A Use of Large Language Models (LLMs): LLM assistance did not substitute for benchmark generation, measurement records, human annotation, or empirical validation, and suggested prose was reviewed against WRBench evidence dimensions.

B Reproducibility and Metric–Dataset Records

The benchmark’s main tables are reproducible because prompts, models, metrics, and aggregation scripts are fixed, while records preserve metadata, exported measurements, analysis slices, and human-validation procedures.

Main tables are generated after fixing the prompt set, model roster, metric definitions, and aggregation scripts.
Benchmark prompts and first frames record initial-state, event, object/action, and camera-trajectory metadata.
The model roster records identifiers, conditioning interfaces, generation settings, and reference-video conditioning for V2V methods.
CamPrec, visual integrity, visible-frame, and re-observed-state consistency metrics, along with re-observation records, are exported before aggregation.
Aggregation scripts generate model, family, camera-protocol, reasoning-tier, and object/action slices, while human validation records sampling, instructions, agreement computation, and calibration examples.

C Limitations and Future Work · D Metric–Dataset Common-Problem Checks

WRBench measures whether viewpoint-conditioned world-state claims remain bound when observability changes, while its diagnostics support model-design improvements and metric-specific interpretation. The benchmark’s next steps include explicit state carriers, axis-level reward or policy signals, finer VLM annotations, dense controls, and denominator-aware analysis.

C Limitations and Future Work: WRBench targets viewpoint-conditioned dynamic world-state attribution under changing observability.Its focus includes whether world-state claims remain bound when camera motion changes what is observable.
C Limitations and Future Work: The benchmark tests whether target-relative displacement, spatial relations, and event endpoints remain bound when observability changes.Natural-25 and the evaluation dimensions make this endpoint-binding problem measurable.
C Limitations and Future Work: Future model-design loops should use these diagnostics to train systems with explicit state carriers.The proposed loop also mines reward or policy signals from axis-level records.
C Limitations and Future Work: The benchmark should be extended with finer VLM-labeled masks, boxes, and dense-control settings.Dense controls are identified as a future extension to the benchmark.
D Metric–Dataset Common-Problem Checks: Table 6 records the intended common-problem interpretation for each WRBench metric.The main text analyzes diagnostic metrics before comparing viewpoint condition types.
D Metric–Dataset Common-Problem Checks: The generated-video dataset is not collapsed into one score because each denominator answers a different question about unsupported world-state claims.Denominator choice therefore distinguishes where a claim becomes unsupported.

D.1 Visual-Integrity Implementation … E.2 Why an in-place change degrades across re-observation

The appendix specifies a DINOv2-based visual-integrity score, expands diagnostic slices and frame-level evidence, and shows that correct appearance or camera access does not guarantee a consistent return. In-place changes fail through distinct manifestations—wrong state, erasure, or target loss—because the unseen endpoint lacks an anchoring coordinate.

D.1 Visual-Integrity Implementation: Visual integrity samples videos at 3 fps, retains the first and last frames, caps clips at 24 frames, and preserves the full field of view through padding.Center cropping is avoided because prompt-critical subjects, targets, or failure evidence may lie outside the center.
D.1 Visual-Integrity Implementation: The visual-integrity score combines global frame-level appearance continuity with local patch-token matching across adjacent sampled frames.Local matching uses valid patches and bidirectional best matches, while the global term uses the DINOv2 CLS token.
D.1 Visual-Integrity Implementation: The 20th-percentile aggregation makes visual integrity sensitive to localized collapse, ghosting, disappearance, hard cuts, and identity drift despite otherwise similar frames.The method tolerates object and camera motion but is not an object detector, tracker, or prompt-grounded subject mask.
D.1 Visual-Integrity Implementation: The CLS-based global descriptor is reported because it outperformed the valid-patch-mean alternative on exact agreement, tie accuracy, and weighted κ in the adjudicated bridge slice.Padding patches are excluded from local matching, although the CLS token can attend to padded regions.
D.2 Additional Common-Problem Tables and Case Montages: Higher re-observation support changes what can be judged but does not automatically produce re-observed consistency; state-only pressure is treated as an event-design effect.The appendix separates re-observation support from re-observed consistency and retains action-level pressure tables and per-sample montages.
E Frame-Level Failure Forensics: The frame-level appendix grounds aggregate findings in raw released outputs by varying exactly one factor per claim and decoding evenly spaced frames.Row labels use the same per-clip D2–D6 values as the aggregate analysis, covering visual quality, in-place changes, camera access, and shared in-place failures.
E.1 Visual quality does not certify a correct return: A clean image, accurate camera move, and returning subject can all occur while the return remains wrong, so these successes do not compose automatically.The cited VerseCrafter example reports D2 = 0.85 and D3 = 0.81 before showing an incorrect return.
E.2 Why an in-place change degrades across re-observation: In-place changes produce wrong-state, erasure, or target-loss failures because the unseen endpoint was never bound to a new coordinate.In the sit example, target loss drives re-observed-spatial to D5 = 0.27.

E.3 Why access depends on the camera channel · E.4 All interfaces fail on the same in-place case

Camera motion determines whether hidden-then-returned events can be tested, but access does not ensure correct re-observation. Across all four interfaces, the in-place folding event fails to return in its correct unobserved endpoint.

E.3 Why access depends on the camera channel: Right-to-left yaw cleanly removes and returns the person, producing 88.5% re-observation support, whereas left-to-right yaw does not.The camera direction changes access while the Gen3C lobby scene remains fixed.
E.3 Why access depends on the camera channel: Prompt-only Kling produces high visual integrity (D2 ≈0.93) but never leaves the subject out of frame, creating no judgeable re-observation.Its bottleneck is failure to create the test, not rendering quality.
E.4 All interfaces fail on the same in-place case: Source-video and geometry-cache interfaces provide access to testing, but that access does not guarantee a correct return.The benchmark distinguishes creating a hidden-then-returned event from preserving its resulting state.
E.4 All interfaces fail on the same in-place case: In the fixed bed-folding task, Gen3C warps the bed and blanket after re-projecting an unobserved fold, yielding D6 = 0.38.The geometry cache never observed the fold.
E.4 All interfaces fail on the same in-place case: ReCamMaster merges two beds during the camera sweep instead of preserving one, yielding D6 = 0.45.This is the source-video interface’s characteristic failure on the same folding event.
E.4 All interfaces fail on the same in-place case: Wan-Fun A14B keeps the cleanest frame (D2 = 0.89) but couples the subject to ego-motion, preventing fold re-observation and yielding D6 = 0.58.The subject is dragged along the camera direction rather than remaining available for the in-place event’s return.
E.4 All interfaces fail on the same in-place case: Kling never lets the subject leave the frame, so no test is created and the correctly folded state is not returned.Thus the unobserved in-place endpoint is preserved by none of the four interfaces.

E.5 High Scores, Frame-Level Failures: The Metric Paradox · F Model Subtype and Series Reading Guide

Aggregate scores can conceal severe frame-level failures, as shown by models that excel in camera execution, re-observation access, or visible quality while failing on returned-state fidelity. The reading guide distinguishes model conditions and assigns text-plus-reference-video methods to the source-video condition rather than a separate family.

E.5 High Scores, Frame-Level Failures: The Metric Paradox: Aggregate dimension scores can rank models highly on one axis while another collapses, making frame-level reading necessary to expose the failure.Headline scores and failures may occupy separate axes.
E.5 High Scores, Frame-Level Failures: The Metric Paradox: D1 = 0.822 makes HyDRA the strongest camera executor, but D6 = 0.445 makes it the weakest re-observed world.HyDRA also records common-yaw alignment of 0.855 and visual integrity D2 = 0.691.
E.5 High Scores, Frame-Level Failures: The Metric Paradox: 73% re-observation support makes Gen3C the leader in access, yet its 3D-cache reprojection produces first-frame ghosting and progressive degradation.The first ∼0.4 s contains a translucent doubled overlay, while leaving the cached frustum leads to smears and darkening.
E.5 High Scores, Frame-Level Failures: The Metric Paradox: Gen3C’s 57.5% and 88.5% support rates show that camera motion determines whether a hidden-then-returned event is created.The L→R sweep keeps the subject framed, whereas the R→L sweep cleanly hides and returns it; the latter has 0.65 retention.
E.5 High Scores, Frame-Level Failures: The Metric Paradox: LiveWorld’s visible spatial score exceeds 0.9, but in-place-change events reduce re-observed state to D6≈0.31–0.40 because its monitor agent hallucinates unseen subjects.Its easy clips reappear cleanly, while the failure concentrates on events requiring an out-of-sight state change.
F Model Subtype and Series Reading Guide: The appendix compares metric–dataset issues, model subtypes, viewpoint conditions, targeted series, human-aligned interpretation, and future reward or policy-training use.It records assignment rules for condition-type and series comparisons.
F Model Subtype and Series Reading Guide: Methods taking text plus a source or reference video are assigned to the source-video condition rather than a separate top-level family.Generated artifacts use V2V, TI2V, and I2V labels.

F.1 Subtype Taxonomy

The subtype taxonomy groups WRBench rows by the dominant viewpoint-condition type that supplies information for the requested viewpoint change, while treating control and state modules as modifiers. Assignment is made per evaluated row and resolves hybrid systems by the evidence path used for re-observation.

Condition types: WRBench’s primary taxonomy variable is the dominant viewpoint condition type supplying information about the requested viewpoint change.Source-video uses reference streams; geometry-cache uses point clouds, 3D caches, or 3D–4D controls; model-inferred uses internally inferred conditions; prompt-only uses prompts.
Assignment rules: Assignment is made at the evaluated WRBench-row level rather than at the model-family level.If a future system exposes multiple interfaces, it is assigned by the interface and evidence path used in the reported WRBench run.
Condition types: The taxonomy includes source-video, geometry-cache, model-inferred, and prompt-only conditions, with control and state modules recorded as modifiers.Representative model-inferred systems include Gen3C, Spatia, VerseCrafter, Wan-Fun, LingBot, LiveWorld, Hunyuan GameCraft, Hunyuan WorldPlay, and MagicWorld; prompt-only rows include Hailuo, Kling, Wan API rows, and HappyHorse.
Assignment rules: For hybrids, external source-video or geometry-cache evidence takes precedence when it supplies view-state information for re-observation; otherwise local modules remain modifiers within model-inferred.Proprietary API rows remain prompt-only unless strict requested-control artifacts are available.
Interpretive use: The series comparisons are diagnostic contrasts rather than scalar rankings, separating visible preservation, camera-target access, access mechanisms, and re-observed consistency.The contrasts include LingBot versus Wan-Fun, Wan scaling, source/memory/geometry/local-control routes, and the Hunyuan split.

F.2 Series Comparison Notes · F.3 Condition-Type Detail Notes · F.4 Directional Access Diagnostic

The appendix compares control series and condition types through access, visible preservation, and re-observed consistency, finding that stronger observability or spatial replay does not establish hidden endpoint binding. Directional performance also requires geometry-calibrated interpretation because benchmark layout can make one yaw direction easier than the other.

F.2 Series Comparison Notes: LingBot preserves visible spatial/state quality, whereas Wan-Fun executes stronger camera displacement and achieves higher re-observation support at lower visible event scores.LingBot World and LingBot Act have only 6.0%/6.4% re-observation support, making their consistency readings sparse.
F.2 Series Comparison Notes: Larger Wan variants improve or preserve visible rendering and raise re-observation support more clearly than conditional re-observed consistency.This within-family diagnostic does not establish a universal capacity rule: larger backbones can expand judgeable re-observed samples without binding hidden event endpoints.
F.2 Series Comparison Notes: State-carrier comparisons show that camera-lens control stores nothing out of sight, static geometry records yield low re-observed state, and source-video carriers provide the highest access.Reported examples include 12.0–18.2% support with 0.62–0.66 re-observed state for camera-lens control, versus ReCamMaster at 58.5%/0.616 and InSpatio at 62.3%/0.664.
F.2 Series Comparison Notes: Training-signal and paradigm comparisons indicate that synthetic paired data can strengthen access and camera alignment, while no listed post-training method supervises the hidden event endpoint.The endpoint gap is therefore described as an unwritten training objective rather than a property of the camera interface.
F.2 Series Comparison Notes: WorldPlay improves access relative to GameCraft because chunk memory is more compatible with re-observation pressure, but scene-context memory still does not establish re-observed consistency.GameCraft’s player-action interface is partly mismatched to WRBench’s prescribed yaw and target-region re-observation protocol.
F.3 Condition-Type Detail Notes: Source-video conditions supply appearance, layout, and temporal evidence through a reference stream, increasing re-observation support without requiring an explicit point-cloud cache.Their dynamics are borrowed from the conditioning stream rather than inferred as a persistent hidden state.
F.3 Condition-Type Detail Notes: Geometry caches make camera-target access computable and dominate high-support rows, while model-inferred and prompt-only conditions can remain visually plausible without reliably producing hidden-then-returned evidence.Gen3C is the access reference, but its re-observed-consistency values are still not saturated; prompt-only rows are outside strict requested-control CamPrec.
F.4 Directional Access Diagnostic: Yaw-direction gaps require a geometry-anchored baseline because benchmark layout can make one direction easier; larger deviations from calibrated retention are the stronger model-failure signal.The diagnostic supports condition-type analysis and hard-case mining, while the main claim remains focused on access.

G Human Validation Notes · H Preference-Pair Export and Reward/Policy Outlook · H.1 Reward and Policy-Training Outlook

The appendices establish human-calibrated validation and a measurement-only handoff for future preference-based reward or policy training. They emphasize that optimization should preserve judgeable hidden-and-returned state evidence rather than reward visible plausibility alone.

G Human Validation Notes: WRBench human comparisons use requested-camera precision, prompt-camera alignment, visual integrity, visible spatial/state correctness, re-observation support, and re-observed consistency.These dimensions provide the calibration framework for the reported diagnostics.
G Human Validation Notes: 2,641 annotation rows produced 2,547 deduplicated annotator verdicts across 1,156 comparison pairs and five semantic families.The released subset and annotation inventory support the human-validation registry and agreement analyses.
G Human Validation Notes: 0.730/0.773/0.702 exact/strict/tie agreement, weighted κ = 0.678, and rank ρ = 0.709 calibrate the DINOv2 visual-integrity proxy on a 190-pair holdout.The visual-integrity proxy is calibrated separately from world-state VLM probes; the world-state bridge uses 230 comparisons over 261 unique videos.
G Human Validation Notes: Visible spatial and visible event-state metrics produce no human-preference-to-opposite reversals, while re-observed spatial and event-state metrics produce one and eight reversals.Re-observed scores are conditional on judgeability, with hidden-and-returned evidence assessed separately from conditional consistency accuracy.
H Preference-Pair Export and Reward/Policy Outlook: WRBench metric records can be exported as preference pairs, but the paper reports no preference optimization, reward-model training, DPO, or policy-improvement results as evidence.The appendix documents the export interface and summarizes counts after missing assets are removed.
H.1 Reward and Policy-Training Outlook: Reward models can mine pairs from WRBench dimensions, while policy-training studies can test whether optimizing them changes held-out evidence dimensions.One proposed preference favors stronger target-relative camera displacement when visual integrity and returned-evidence judgeability are preserved.
H.1 Reward and Policy-Training Outlook: The current handoff supplies metric records, judgeability flags, pair-construction rules, and dynamic-versus-dynamic filtering for downstream reward and policy experiments.Rewarding visible plausibility alone could reinforce keeping the salient subject in frame without creating a judgeable hidden-and-returned world-state test.

H.2 Dense-Control and Future Benchmark Extensions

Future WRBench extensions should add dense-control models only under leakage-aware input policies, because endpoint-encoding controls can measure control following rather than hidden-state inference. Evaluation should separate source-only, endpoint-masked, and full-target controls, while training claims require held-out evidence.

Dense-control settings: Dense-control systems should enter WRBench as future model clusters with explicit input-policy labels to prevent endpoint leakage from masquerading as hidden-state inference.Target-view depth, segmentation, edge maps, LiDAR, or HD maps may already encode the post-event endpoint.
Dense-control settings: Source-only control extracts inputs from the source view or video, withholding the target-view endpoint.This setting tests the claim without supplying the endpoint through target-view controls.
Dense-control settings: Endpoint-masked target control permits target-view controls while masking or ambiguating the object/contact endpoint region.The endpoint remains unavailable or unclear despite allowing target-view control information.
Dense-control settings: Full target control provides an upper-bound control-following condition, not evidence that the model maintains internal world-state persistence.Target-view dense controls can supply the endpoint directly, so performance should not be interpreted as hidden-state inference.
Training and evaluation: Reward or policy-training gains require held-out evidence, although WRBench records can mine preference pairs over camera displacement, visual integrity, re-observation, and consistency.The preference dimensions are target-relative camera displacement, visual integrity, judgeable re-observation, and re-observed consistency.