Yuxue Yang
Shuyao Shang
Jiahe Wang
Zitong Zhou
Liang Tan
Junhan Zeng
Ruizhi Li
Junyan Li
Yu Liu
Xiao Yang
Yong Li
Tieniu Tan
Lue Fan
Zhaoxiang Zhang
Abstract
Controllable video generation models are increasingly being developed as world models. Accordingly, evaluating them in this role extends beyond the apparent appearance of generated videos to the inherent reactivity of the worlds they depict: the ability to infer from the scene state how the world should react and to generate plausible consequences not explicitly described in the input. Yet existing benchmarks mainly assess visual quality or explicit instruction fulfillment by checking whether requested actions and interaction outcomes are realized, leaving inherent reactivity underexamined. We introduce WorldExam, a hierarchical diagnostic benchmark spanning four levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. It comprises 1,474 cases across eight dedicated tasks and supports unified evaluation of camera-, action-, and language-driven model paradigms. The World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input. Evaluation of 20 representative models reveals a clear capability split. Camera-driven models excel at camera control, but their interfaces do not support dynamic interaction; action-driven models control subjects more precisely but often leave the world unresponsive; and language-driven models perform better on interaction but follow complex controls less faithfully. No model combines broad task coverage with consistently strong performance, showing that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.
1 Introduction
Controllable video generation models are increasingly being developed as world models rather than standalone clip generators [brooks2024video, bruce2024genie, yang2023unisim, recammaster, neoverse, worldplay, lingbot, vidu, veo]. Such models are expected to predict future visual states from an initial observation and control instructions, including camera trajectories, action sequences, and language prompts. Evaluating them in this role extends beyond the apparent appearance of generated videos to the inherent reactivity of the worlds they depict [yang2026mirabench, li2026robotrustbench]. When a subject moves onto stairs, its motion should adapt to the terrain; when it approaches an obstacle, the world should show contact, avoidance, or blockage; when it enters another agent’s personal space, that agent should respond plausibly. These effects are scene-conditioned consequences rather than direct depictions of the input. Together, they reveal a model’s inherent reactivity: its ability to infer from the scene state how the world should react and to generate such consequences plausibly.
Recent benchmarks have advanced world-model evaluation beyond perceptual quality to structured layout control [duan2025worldscore], unified action interfaces [ye2026mind, xu2026worldmark, fang2026iworld, ying2026wbench, xu2026worldroambench], prompt-specified interaction effects [wu2026omniworldbench, zhao2026worldolympiad], and embodied-AI and autonomous-driving applications [shang2026worldarena, liang2025worldlens]. As summarized in table˜1, they span camera-, action-, and language-driven model paradigms and increasingly cover camera and subject control, scene revisiting, and interaction outcomes. Complementary benchmarks probe implicit rules, future-state reasoning, and law-specific physical consistency in specialized settings [liu2026risevideo, wu2026worldreasonbench, upadhyay2026worldbench, lin2026phyground]. Yet most benchmarks still assess explicit instruction fulfillment: a desired layout, camera trajectory, action sequence, or interaction consequence is specified in advance, and the model is evaluated on whether the specified outcome is realized. This evaluation is necessary, but it leaves underexamined a model’s ability to infer additional consequences implied by the initial state but not described in the instruction.
We introduce WorldExam, a hierarchical diagnostic benchmark designed around this distinction, as summarized in figure˜1. WorldExam represents each controllable behavior as a composition of atomic control units and adapts these units to each model’s native interface: camera trajectories for camera-driven models, discrete action sequences for action-driven models, and natural-language prompts for language-driven models. For World Reactivity cases, the model-facing instruction specifies only the explicit control or goal, leaving the expected scene-conditioned reactions unstated. This design distinguishes direct fulfillment of a requested outcome from behavior beyond what the input explicitly specifies.
WorldExam organizes evaluation into the four diagnostic levels: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. We instantiate this hierarchy with eight evaluation tasks: Camera Control, Subject Control, Scene Revisit, Terrain Interaction, Object Interaction, Social Interaction, Physical Reaction, and Goal Completion. Visual Quality is measured with task-agnostic metrics; Control Adherence is evaluated through Camera Control and Subject Control; and Spatial Consistency is evaluated through Scene Revisit. The World Reactivity level covers scene-conditioned reactions and goal-directed behaviors. Within this level, four reaction-oriented tasks use control units as triggers while leaving the induced scene-conditioned reactions unstated. Goal Completion extends the same principle to goal-directed behavior: it specifies a high-level goal while leaving the detailed execution steps unstated. For example, a goal to arrange three bolts by height specifies the target layout, but not which object to move first or how to realize the motion frame by frame.
For model-interface compatibility, WorldExam uses two tracks rather than one global ranking. The static-scene track controls only the camera and is available to all three paradigms. The dynamic-interaction track requires observable subject–environment interaction and is therefore evaluated only on compatible action- and language-driven models. Separating the tracks avoids treating unsupported capabilities as failures or averaging scores obtained under different scene assumptions and task sets.
Our evaluation of 20 representative models reveals clear trade-offs across the four levels and three paradigms. Camera-driven models excel at camera control, but their interfaces do not support dynamic interaction. Action-driven models control subjects more precisely but often leave the world unresponsive. Language-driven models perform better on interaction tasks but follow complex controls less faithfully. These capability splits are obscured by aggregate scores, motivating separate reporting at both level and task granularity.
- •
We extend world model evaluation beyond apparent appearance to inherent reactivity: inferring from the scene state how the world should react and generating plausible consequences absent from the input.
- •
We propose WorldExam, a benchmark of 1,474 cases across eight tasks that supports unified evaluation of camera-, action-, and language-driven model paradigms.
- •
We evaluate 20 representative models, revealing paradigm-dependent capability splits. No model combines broad task coverage with consistently strong performance, showing that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity.
- •
We will publicly release the benchmark data and evaluation toolkit to facilitate systematic evaluation and foster continued progress in the video world model community.
| Benchmark | Model Paradigm | Viewpoint | WorldExam Evaluation Tasks | #Cases | #Models | ||||||||
| First Person | Third Person | Camera Control | Subject Control | Scene Revisit | Terrain Inter. | Object Inter. | Social Inter. | Physical React. | Goal Compl. | ||||
| WorldScore [duan2025worldscore] | C/L | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | 3,000 | 20 |
| MIND [ye2026mind] | A | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | 250 | 2 |
| Omni-WorldBench [wu2026omniworldbench] | C/L | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | 1,068 | 18 |
| WorldMark [xu2026worldmark] | C/A/L | ✓ | ✓ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | ✗ | 500 | 6 |
| iWorld-Bench [fang2026iworld] | C/A/L | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✗ | ✗ | 4,900 | 14 |
| WBench [ying2026wbench] | C/A/L | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | 289 | 20 |
| WorldOlympiad [zhao2026worldolympiad] | A/L | ✓ | ✗ | ✓ | ✗ | ✗ | ✗ | ✓ | ✗ | ✓ | ✗ | 1,000 | 8 |
| WorldRoamBench [xu2026worldroambench] | A | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ | 600 | 10 |
| WorldExam (Ours) | C/A/L | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | 1,474 | 20 |
2 Related Work
2.1 Video World Models
Recent video world models increasingly support controllable video generation for gaming, robotics, embodied AI, and open-world simulation. Based on their primary control interfaces, they can be broadly grouped into camera-, action-, and language-driven paradigms. Camera-driven models [trajectorycrafter, recammaster, voyager, fantasyworld, neoverse, inspatio] condition generation on camera trajectories, represented in two main ways. Some approaches, such as ReCamMaster [recammaster] and FantasyWorld [fantasyworld], inject camera trajectories through learned camera encoders or embeddings, whereas others [trajectorycrafter, voyager, neoverse, inspatio] reconstruct 3D priors from the input and reproject them to target viewpoints; representative methods include NeoVerse [neoverse] and InSpatio-World [inspatio]. Action-driven models [gamecraft, astra, worldplay, yume15, lingbot, infiniteworld, matrixgame3] generate future frames conditioned on discrete action sequences through keyboard-like interfaces. Among them, WorldPlay [worldplay] and LingBot-World [lingbot] focus on real-time interaction and consistent generation under direct action control. Language-driven models [kling, veo, hailuo, wan, seedance, vidu, happyhorse] generate videos from text or image-text prompts, demonstrating advances in semantically complex video generation. Across paradigms, video world models are evolving from short open-loop synthesis toward controllable, persistent, and interactive environment simulation. Heterogeneous interfaces complicate direct comparison, while controllability, long-term memory, and inherent reactivity remain key challenges.
2.2 Video World Model Benchmarks
A growing body of benchmarks evaluates complementary aspects of video world modeling. Some emphasize perceptual and temporal quality [huang2024vbench, huang2025vbenchpp, zheng2025vbench2, liu2024evalcrafter, liu2023fetv]; others target compositionality, world knowledge, implicit rules, and future-state reasoning [sun2024t2vcompbench, chen2025t2vworldbench, liu2026risevideo, wu2026worldreasonbench]. Physics-oriented benchmarks [bansal2024videophy, meng2024phygenbench, li2025worldmodelbench, upadhyay2026worldbench, lin2026phyground, xue2026acwmphys, wu2026pdibench] diagnose law-specific dynamics, geometric consistency, and generalization under physical interactions; embodied benchmarks [qin2024worldsimbench, yue2025ewmbench, li2025worldeval, shang2026worldarena, jiang2026robowmbench, yang2026mirabench, li2026robotrustbench, liu2026kinebench] evaluate action fidelity, physical executability, planning utility, reliability, and trustworthiness; and autonomous-driving benchmarks [arai2024actbench, liang2025worldlens, zhou2026drivinggen] emphasize ego-action control, trajectory plausibility, safety, and downstream driving utility. Beyond these settings, general benchmarks [duan2025worldscore, ye2026mind, wu2026omniworldbench, xu2026worldmark, fang2026iworld, ying2026wbench, zhao2026worldolympiad, xu2026worldroambench, zhang2025worldinworld] evaluate interactive world models across varied scenes and interfaces.
Among these general benchmarks, WorldScore [duan2025worldscore] evaluates camera-trajectory-based layout control and geometric consistency, while MIND [ye2026mind] focuses on action control and closed-loop revisit consistency. WorldMark [xu2026worldmark] and iWorld-Bench [fang2026iworld] improve cross-model comparison through standardized or unified action representations. Omni-WorldBench [wu2026omniworldbench] evaluates prompt-specified interaction outcomes, affected and unaffected entities, and intermediate causal state transitions; WBench [ying2026wbench] extends evaluation to multi-turn navigation, subject actions, event editing, and perspective switching; and WorldOlympiad [zhao2026worldolympiad] probes long-horizon interaction and physics. WorldRoamBench [xu2026worldroambench] further couples long-horizon action-conditioned generation with diagnostics of controllability, visual drift, mechanics, optics, 3D consistency, and memory. Collectively, these benchmarks substantially broaden interactive evaluation, but most still center on explicit instruction fulfillment by checking whether specified controls or interaction outcomes are realized. In contrast, WorldExam adapts atomic control units to each model’s native interface and evaluates inherent reactivity through scene-conditioned reactions and goal-directed behaviors beyond what the input explicitly specifies.
3 WorldExam
WorldExam supports unified evaluation of video world models with different control interfaces. We formulate a video world model as a function , where is the initial image, is the model-facing input instruction and is the generated video. We consider three common paradigms: camera-driven models take camera trajectories in , action-driven models take discrete action sequences over {W (move forward), S (move backward), A (move left), D (move right), (tilt up), (tilt down), (pan left), (pan right), (stop)}, and language-driven models take natural-language prompts.
To compare these paradigms, WorldExam uses interface adaptation to map a shared case to each model’s native interface. WorldExam represents controllable behavior as an ordered composition of atomic control units, such as moving forward (W) and then panning right (), and adapts this control intent into an camera trajectory, a discrete action sequence, or a natural-language prompt. Under this setup, WorldExam first defines a four-level diagnostic hierarchy (section˜3.1) and then instantiates it through eight evaluation tasks (section˜3.2). We next describe the test case curation pipeline (section˜3.3) and the benchmark statistics (section˜3.4). Metric definitions and scoring protocols are described in section˜4.
3.1 Toward World Reactivity: A Four-Level Diagnostic Hierarchy
WorldExam organizes world-model evaluation into four diagnostic levels of increasing scope: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. Visual Quality measures the video’s apparent appearance, including perceptual plausibility, temporal stability, and aesthetic quality. Control Adherence measures whether the controlled camera or subject follows the input control. Spatial Consistency measures whether the model preserves a coherent world when the camera revisits a previously observed viewpoint.
By contrast, the World Reactivity level evaluates scene-conditioned reactions and goal-directed behaviors beyond what is explicitly specified in the input. In reaction-oriented cases, a control specifies the initiating behavior but leaves its scene-conditioned consequences unstated, which the model must infer from the scene. For example, a move-forward control specifies the subject’s direction but not how its motion should adapt to the terrain. If there is an obstacle or a nearby agent in its motion path, the subject may stop or avoid it, another agent may yield, or an object may move on contact. In goal-directed cases, a high-level goal specifies the desired target, and the model needs to infer from the initial scene how to realize it frame by frame.
Although the four levels form a diagnostic progression, strong Visual Quality, Control Adherence, and Spatial Consistency do not guarantee successful scene-conditioned reactions or goal execution. Conversely, success on World Reactivity does not compensate for visual artifacts, control errors, or spatial drift. WorldExam therefore reports the four levels separately to localize failures in generation quality, explicit control, spatial persistence, and behavior that must be inferred from the scene.
3.2 From Diagnostic Levels to Evaluation Tasks
Figure˜2 shows how the four diagnostic levels are instantiated. Visual Quality uses task-agnostic metrics across all videos, while the other three levels are instantiated by eight tasks. Control Adherence includes Camera Control and Subject Control, while Spatial Consistency uses Scene Revisit. World Reactivity includes four reaction-oriented tasks, Terrain Interaction, Object Interaction, Social Interaction, and Physical Reaction, plus Goal Completion for goal-directed execution.
The tasks are reported through two tracks rather than a single global score to avoid penalizing models for tasks their interfaces do not support. The static-scene track contains Camera Control and Scene Revisit and is available to all three model paradigms because each interface can express camera motion. The dynamic-interaction track contains Subject Control and the five World Reactivity tasks and applies only to compatible action- and language-driven models; Goal Completion is language-only.
Control Adherence.
Camera Control tests whether the generated camera motion follows the prescribed controls. Each case composes one to three atomic camera controls from {W, S, A, D, , , , }, assigns each an execution-time fraction, and executes them in order over the assigned intervals. Subject Control applies the same construction to a designated third-person subject using {W, S, A, D}.
Spatial Consistency.
Scene Revisit tests the model’s spatial memory of the initial observation. Each case uses a round-trip camera trajectory formed by an outgoing control and its inverse, such as “move left” followed by “move right”, or “tilt up” followed by “tilt down”. After moving away, the camera should return to the initial viewpoint while the returned view preserves the scene’s geometry, appearance, and content.
World Reactivity.
The four reaction-oriented tasks pair an initial scene with a single atomic subject control. Terrain Interaction places stairs, slopes, bridges, trenches, or other structured terrain along the controlled subject’s path. The input specifies only the horizontal motion direction, while the model must infer how the subject should adapt its height and maintain contact with the terrain. Object Interaction places a movable, flexible, or rigid target along the subject’s path so that the subject, one of its body parts, or a carried tool is expected to make contact with it. It evaluates whether the target produces an immediate type-appropriate response, such as motion when loose, deformation when flexible, or blockage when rigid, without interpenetration. Social Interaction places other agents along the subject’s path or within its social distance, creating an imminent local conflict. It evaluates whether the affected agents respond plausibly through avoidance, yielding, stopping, or changing path. Physical Reaction tests whether a dynamic process unfolds over time according to physical regularities, including gravity, friction, momentum transfer, constrained motion, fluid response, and pendulum-like swinging. Each case uses one control from {W, S, A, D, (stop)}. A motion control may trigger the process, whereas is used when the initial scene is expected to evolve autonomously without subject motion. Although Object Interaction and Physical Reaction may both involve contact, the former targets the immediate type-conditioned response of a designated object, whereas the latter targets the temporal evolution of a physical process. Goal Completion is language-only and provides a high-level goal together with an initial scene containing relevant entities, distractors, preconditions, and constraints. Unlike the four reaction-oriented tasks, it uses no atomic control sequence or execution-time fractions. The input may state necessary subgoals or ordering constraints. The model should ground the goal in the initial scene, select the correct entities, ignore distractors, and produce coherent execution steps toward the desired target frame by frame. Section˜9 provides examples of all eight tasks and representative checklists.
3.3 Test Case Curation Pipeline
WorldExam constructs cases differently for the two tracks. For static-scene Camera Control and Scene Revisit, we pair suitable first-person scenes from existing datasets [Flickr2K, dl3dv, ye2026mind] with compositions of atomic control units. For dynamic-interaction Subject Control and the five World Reactivity tasks, the pipeline in figure˜3 constructs initial scenes supporting the intended behavior and evaluation.
For each dynamic-interaction task, a task-specific pattern library defines the intended semantic coverage over subject motion, structured terrain, object contact, social conflict, physical processes, or goal-directed situations. After a pattern is sampled, a schema-guided LLM case composer expands it into a structured draft containing a detailed scene description, an initial-image generation prompt, a control intent or high-level goal, and a draft text prompt for language-driven models. For Object Interaction, Social Interaction, Physical Reaction, and Goal Completion, the draft also includes a case-specific checklist of observable evaluation criteria. In each World Reactivity case, the model-facing input specifies only the explicit control or high-level goal; the scene-conditioned reaction or detailed execution process remains unstated.
The initial-image generation prompt is used only to synthesize candidate initial images. Human filtering retains candidates in which the relevant entities are visible, the spatial layout supports the intended behavior or event, the image is consistent with the draft, and sufficient motion space remains for the continuation. Candidates that already depict the evaluated event or desired target, hide relevant entities, or make the intended behavior physically infeasible are discarded. The selected initial image therefore provides a concrete pre-event state from which the intended behavior, reaction, or goal-directed execution can unfold.
An image-conditioned case refiner then revises the draft to match while preserving the sampled pattern and intended control or goal. It updates entity references, spatial relations, the scene description, the text prompt, and the optional checklist so that all referenced entities and preconditions are grounded in the selected image. For tasks evaluated using checklists, the final checklist is fixed at this stage. The finalized case consists of , the control intent or high-level goal, the grounded text prompt, the optional checklist.
3.4 Benchmark Statistics
As shown in figure˜2, WorldExam contains 1,474 cases across eight evaluation tasks. Figure˜4 summarizes both the overall dataset composition and the task-specific composition of the five World Reactivity tasks.
At the dataset level, the cases span first- and third-person viewpoints with first-person viewpoints accounting for 31.4% of the benchmark and providing substantial egocentric coverage. The subject taxonomy spans humans, animals, vehicles, and robots, while the visual-style taxonomy mixes outdoor and indoor real scenes with 3D renderings, cinematic footage, close-up views, animation, and dashcam videos. Scene content is also deliberately broad: no single scene type dominates the benchmark, and the largest category, traffic scenes, accounts for only 14.7% of the cases.
Within the five World Reactivity tasks, the cases are further distributed across task-specific semantic subcategories, including terrain types, social-interaction scenarios, object types, physical-reaction types, and Goal Completion domains and task types. No single subcategory accounts for more than 35% of its corresponding task. This coverage reduces dependence on any one visual or semantic template and supports task-specific analysis across diverse scene-conditioned reactions and goal-directed situations. Representative cases across these dimensions are shown in section˜7.
4 Evaluation Protocol and Metrics
For Camera Control, Scene Revisit, Subject Control, and Terrain Interaction, we lift each generated video into 3D with a geometry reconstruction model and evaluate it in the reconstructed space. Camera Control and Scene Revisit use the recovered camera trajectories, whereas Subject Control and Terrain Interaction use the recovered 3D subject trajectories and terrain geometry. For the remaining four tasks, we use GPT-5.5 as the vision-language model (VLM) judge to score generated videos against predefined case-specific checklists. Each track reports task-specific metrics together with task-agnostic general metrics for visual quality.
4.1 Static-Scene Track
Given generated frames , we use VGGT- [wang2026vggt] to estimate camera poses, intrinsics, and depths.
Camera Control.
Each case specifies an ordered sequence . Here, {W, S, A, D, , , , } is an atomic control unit, and is its execution-time fraction, with . For camera- and action-driven interfaces, we allocate frames to the -th control using nearest-integer rounding, and adjust the allocation to ensure . From each generated video, we recover a frame-wise camera trajectory . Because the three interfaces specify camera motion differently, we construct the model-facing input and evaluation reference separately for each interface.
For camera-driven models, controls are composed sequentially starting from the initial camera pose, with each subsequent control applied relative to the endpoint pose of the previous control. These endpoints serve as keyframes, which we interpolate over the allocated frame intervals to obtain a frame-wise input trajectory in . This input trajectory also serves directly as the frame-wise reference trajectory . Before comparison, we express both the recovered and reference trajectories relative to their respective first-frame poses. We then compute the translation and rotation errors between the two trajectories as
| (1) |
The nonnegative scale resolves the translation-scale ambiguity of monocular camera reconstruction, making scale-invariant, while converts from radians to degrees. In implementation, the argument of is clipped to for numerical stability.
Action-driven models map each control to the model’s native discrete action and assign the corresponding frames to the -th action. Language-driven models instead verbalize each control, join the resulting motion phrases in order with “then,” and prepend the instruction to the scene description; for example, “W” followed by “” becomes “The camera moves forward, then pans right. [Scene description].” This prompt preserves the control order but does not specify the duration of each control.
Unlike camera-driven interfaces, action- and language-driven interfaces do not specify an exact camera trajectory in , so we evaluate their recovered trajectories against control-level references segment by segment. For action-driven models, the frame ranges assigned to the discrete actions directly define the segment boundaries. For language-driven models, we instead partition the recovered trajectory into segments by applying dynamic-programming-based change-point detection [ruptures] to frame-to-frame changes in translation and rotation. The resulting segments are matched in temporal order to the atomic controls.
Within each segment, we express the recovered camera poses relative to the first frame, so that the segment starts from the identity pose. The assigned control determines whether the reference motion is a translation or a rotation. For a translation control, we linearly interpolate the reference translation from to a unit vector in the prescribed direction, while keeping throughout. The unit displacement is sufficient because is invariant to translation scale. For a rotation control, we set and construct using the prescribed axis and direction; its translation error is computed as the mean per-frame without scale alignment. Because the interface does not specify a rotation angle, the reference angle is linearly interpolated from zero to the total angle recovered within the segment. The segment-level references therefore evaluate the prescribed direction and motion progression without imposing a fixed magnitude.
For each segment , we compute using the error definitions above. A translation segment is assigned the maximum translation error if its displacement is below 5% of the largest segment displacement in the same video or if its net motion is not aligned with the prescribed direction. A rotation segment is assigned the maximum rotation error if its total rotation angle is below or is opposite to the prescribed direction. Finally, we obtain the video-level errors by averaging the corresponding segment-level errors using the number of frames in each segment as weights.
Across all interfaces, we normalize the resulting errors as and , where is measured in degrees, and report their geometric mean, . Thus, a high Camera Control score requires the recovered camera motion to follow the prescribed directions and temporal progression.
Scene Revisit.
Scene Revisit evaluates two requirements after a round-trip camera motion: returning the camera to its initial pose and preserving the initial scene in the returned view. Each case pairs an outgoing control with its inverse . We set their execution-time fractions to and , reserving a longer temporal window for the return motion so that the model has sufficient opportunity to reach the initial viewpoint. We adapt this pair to the three interfaces as in Camera Control.
Let denote the recovered camera pose. For camera- and action-driven models, is the frame range allocated to ; for language-driven models, whose prompt does not specify control duration, it begins at 40% of the video. Because the camera may return before the video ends, we search this entire segment and select the frame whose recovered pose is closest to the initial pose:
| (2) |
For translation round trips, ; for rotation round trips, is the relative rotation angle between and . A translation revisit succeeds when this minimum distance is within 10% of the maximum displacement reached during the outgoing segment; a rotation revisit succeeds when its minimum angular distance is below . Averaging this binary result over all cases gives Revisit Success .
We then compare the input image with the selected revisit frame using PSNR, LPIPS, and SSIM. Selecting the frame by recovered pose rather than using the final frame makes this appearance comparison insensitive to small differences in return timing. After averaging over cases, we normalize the three appearance metrics as , , and , and combine them with Revisit Success:
| (3) |
The geometric aggregation gives a high Scene Revisit score only when the camera both returns to the initial viewpoint and recovers a consistent view.
General metrics.
Across all static-scene videos, we additionally report five task-agnostic general metrics. 3D Consistency adapts the metric of WorldScore [duan2025worldscore] to VGGT- outputs. Using the recovered geometry, we project valid pixels from a source frame to a nearby target and back, then measure the cycle reprojection error. Photometric Consistency measures the forward-backward optical-flow cycle error between neighboring frames using average endpoint error (AEPE). Temporal Flickering, Aesthetic Quality, and Imaging Quality are adapted from VBench [huang2024vbench]. These metrics summarize whether the static-scene generations are geometrically stable, temporally coherent, and visually plausible.
4.2 Image-Space Displacement Alignment for Camera-Driven Models
Identical translation values in an input trajectory can induce different image-space displacements across camera-driven models because their pose-conditioning interfaces interpret the translation magnitude differently. Larger displacements expose more novel-view content and increase the difficulty of generating controlled, spatially consistent videos. By increasing the input translation multiplier to induce progressively larger image-space displacements, the results in tables˜3 and 7 show corresponding declines in Camera Control, Scene Revisit, and general-metric performance.
The scale-invariant translation error in section˜4.1 does not address this difference. Its scalar is fitted after generation only to resolve the coordinate-scale mismatch between the recovered and reference trajectories; it neither changes the input trajectory nor normalizes the image-space displacement in the generated video. We therefore calibrate input translations before generation to align image-space displacement (figure˜5).
For each model and case , we estimate a translation calibration factor using the anchor mask provided on the initial frame. We first generate a calibration video with the model’s default input translation magnitude, apply a horizontal camera control, either “move left” or “move right”, and track the anchor through the video using SAM2 [sam2]. Let denote the horizontal image-space displacement of the tracked mask center, and let denote the frame width. We set the target displacement to and compute . For the final generation used in evaluation, we multiply the translation components of the default input trajectory by while leaving its rotations unchanged.
4.3 Dynamic-Interaction Track
Subject Control.
Subject Control uses the same ordered-control construction as Camera Control, with {W, S, A, D} applied to a designated subject. Action-driven models receive native discrete subject actions over the allocated frame ranges. For language-driven models, we verbalize the ordered controls together with the designated subject and prepend the resulting instruction to the scene description; for example, “W” followed by “A” becomes “The [subject] moves forward, then the [subject] moves left. [Scene description].”
As illustrated in figure˜6, SAM2 [sam2] tracks the designated subject from a first-frame mask, and VGGT- [wang2026vggt] lifts the tracked pixels into 3D to recover the subject trajectory. The horizontal-region mask estimates gravity and the horizontal plane; projecting the initial camera’s viewing direction onto this plane defines the forward reference direction, with the other directions derived analogously.
We evaluate the recovered subject trajectory segment by segment. The action-driven segment boundaries follow the frame ranges assigned to the discrete subject actions, whereas the language-driven segments are inferred by applying the same change-point procedure as in Camera Control to frame-to-frame subject displacement. Within each segment, we translate the recovered trajectory so that its first-frame subject position is the origin. The associated atomic control unit selects a horizontal unit direction , and the reference trajectory is linearly interpolated from the origin to . We fit a post-generation translation scale between the recovered and reference trajectories, as in Camera Control, and compute the segment-level translation error . A segment whose net displacement is below 0.5% of the reconstructed scene scale or whose motion is misaligned with the prescribed direction receives the maximum error . Finally, the error is obtained by frame-count-weighted averaging, and the Subject Control score is .
Terrain Interaction.
Unlike Subject Control, Terrain Interaction uses a single atomic control unit to induce a subject-terrain interaction. For evaluation, we additionally provide a subject-free terrain image to recover the complete terrain geometry. We then project the recovered 3D subject trajectory along gravity onto this geometry to obtain the corresponding 3D terrain trajectory.
During evaluation, we first compute the Subject Control score from the horizontal component of the subject trajectory and use it as a gating check; cases that fail this check are considered not to follow the control and receive a Terrain Interaction score of zero. For cases that pass this check, we extract local extrema and the endpoint from the height of the terrain trajectory as evaluation points. The Terrain Interaction score is the ratio of the number of evaluation points at which the subject and terrain trajectories exhibit consistent height changes to the total number of evaluation points.
Checklist-Based World Reactivity Evaluation.
Considering that the remaining four tasks require semantic and causal judgments that cannot be captured by recovered trajectories, we evaluate them with a VLM judge against the case-specific checklist constructed in section˜3.3. Each checklist covers the initiating condition, the resulting reaction or goal execution progress, and invalid outcomes. At evaluation time, the VLM receives 10 temporally ordered frames uniformly sampled from the generated video, together with the checklist. Using a task-specific prompt, the judge returns one binary decision per item; contradicted, missing, ambiguous, off-screen, or otherwise unverifiable evidence is counted as unsatisfied. The case score is
| (4) |
Object Interaction.
The checklist verifies that contact occurs with the designated object, precedes and causes the reaction, and produces a type-consistent outcome without interpenetration. It also checks that the direction and extent of the reaction remain consistent with the contact.
Social Interaction.
The checklist verifies that the controlled motion creates the intended conflict and that at least one visible affected agent makes a timely adjustment attributable to the controlled subject. Unchanged, delayed, unrelated, or physically implausible responses are counted as failures.
Physical Reaction.
The checklist verifies the timing and cause of the process, its evolution under the relevant physical regularity, and the preservation of required supports, attachments, contacts, and constraints. Freezing, premature onset, interpenetration, broken attachments, or unexplained energy are counted as failures.
Goal Completion.
The checklist separately evaluates correct grounding, intermediate progress, compliance with stated ordering and scene-dependent constraints, and final completion. Scores credit partial progress and accept alternative executions that reach the desired target under the same observable requirements.
General metrics.
The dynamic-interaction track separately reports four VBench metrics [huang2024vbench]: Subject Consistency and Motion Smoothness for feature and temporal consistency, and Aesthetic Quality and Imaging Quality for visual appeal and frame-level quality. We omit 3D Consistency, Photometric Consistency, and Temporal Flickering because valid motion and state changes disrupt their geometric and optical-flow correspondences.
5 Experiments
| Model | Camera Control | Scene Revisit | 3D Cons. | Photo. Cons. | Temp. Flick. | Aesth. Quality | Imag. Quality | Average | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T. Err. | R. Err. | Score | Success | PSNR | LPIPS | SSIM | Score | Task | General | Overall | ||||||
| Camera-driven | ||||||||||||||||
| TrajectoryCrafter [trajectorycrafter] | 0.14 | 1.01 | 80.32 | 1.000 | 19.08 | 0.240 | 0.545 | 83.03 | 96.98 | 62.83 | 92.56 | 52.17 | 67.03 | 81.68 | 74.31 | 78.00 |
| ReCamMaster [recammaster] | 0.47 | 3.91 | 38.64 | 0.815 | 16.02 | 0.353 | 0.415 | 68.01 | 99.33 | 81.97 | 95.52 | 54.58 | 73.45 | 53.33 | 80.97 | 67.15 |
| Voyager [voyager] | 0.31 | 4.29 | 56.19 | 0.995 | 16.88 | 0.388 | 0.466 | 76.27 | 89.54 | 27.43 | 93.04 | 53.59 | 63.69 | 66.23 | 65.46 | 65.84 |
| FantasyWorld [fantasyworld] | 1.04 | 7.82 | 18.46 | 0.560 | 15.41 | 0.323 | 0.374 | 55.79 | 98.57 | 75.60 | 95.91 | 57.11 | 73.98 | 37.12 | 80.23 | 58.68 |
| NeoVerse [neoverse] | 0.01 | 0.60 | 97.33 | 1.000 | 21.72 | 0.141 | 0.662 | 89.25 | 98.19 | 70.83 | 93.53 | 52.72 | 72.19 | 93.29 | 77.49 | 85.39 |
| InSpatio-World (1.3B) [inspatio] | 0.09 | 1.24 | 85.94 | 1.000 | 20.79 | 0.224 | 0.606 | 85.90 | 98.44 | 64.88 | 93.68 | 53.78 | 73.60 | 85.92 | 76.88 | 81.40 |
| Action-driven | ||||||||||||||||
| Hunyuan-GameCraft [gamecraft] | 0.09 | 11.95 | 41.33 | 0.385 | 13.36 | 0.542 | 0.367 | 41.77 | 93.61 | 53.93 | 93.91 | 54.97 | 71.69 | 41.55 | 73.62 | 57.59 |
| Astra [astra] | 0.31 | 7.92 | 32.59 | 0.365 | 12.20 | 0.601 | 0.309 | 38.15 | 96.46 | 78.41 | 96.38 | 51.71 | 71.78 | 35.37 | 78.95 | 57.16 |
| WorldPlay [worldplay] | 0.04 | 1.04 | 92.74 | 0.790 | 18.42 | 0.271 | 0.531 | 72.51 | 98.69 | 81.06 | 95.90 | 53.92 | 73.31 | 82.63 | 80.58 | 81.61 |
| Yume 1.5 [yume15] | 0.12 | 3.23 | 75.67 | 0.255 | 12.37 | 0.615 | 0.359 | 32.45 | 98.44 | 67.53 | 95.18 | 53.42 | 75.28 | 54.06 | 77.97 | 66.02 |
| LingBot-World [lingbot] | 0.11 | 6.22 | 58.19 | 0.605 | 15.58 | 0.341 | 0.406 | 58.35 | 99.59 | 81.52 | 96.53 | 59.87 | 75.00 | 58.27 | 82.50 | 70.39 |
| Infinite-World [infiniteworld] | 0.17 | 2.35 | 71.70 | 0.645 | 14.31 | 0.388 | 0.345 | 57.34 | 99.91 | 92.09 | 95.99 | 56.27 | 76.41 | 64.52 | 84.13 | 74.33 |
| Matrix-Game 3.0 [matrixgame3] | 0.06 | 4.36 | 76.36 | 0.860 | 13.82 | 0.485 | 0.372 | 64.25 | 95.59 | 30.14 | 93.40 | 48.80 | 73.39 | 70.30 | 68.26 | 69.28 |
| Language-driven | ||||||||||||||||
| Kling 2.5 [kling] | 0.19 | 6.90 | 50.18 | 0.325 | 13.54 | 0.532 | 0.381 | 38.81 | 99.87 | 87.87 | 97.56 | 56.28 | 75.70 | 44.50 | 83.46 | 63.98 |
| Veo 3.1 [veo] | 0.27 | 8.22 | 40.83 | 0.447 | 12.30 | 0.594 | 0.345 | 43.05 | 99.06 | 73.61 | 95.43 | 55.49 | 76.57 | 41.94 | 80.03 | 60.99 |
| Hailuo 2.3 [hailuo] | 0.15 | 5.20 | 63.29 | 0.505 | 13.52 | 0.519 | 0.387 | 48.70 | 99.38 | 79.88 | 95.33 | 56.30 | 75.52 | 55.99 | 81.28 | 68.64 |
| Wan 2.6 I2V [wan] | 0.17 | 6.12 | 57.72 | 0.495 | 13.28 | 0.527 | 0.353 | 47.32 | 99.49 | 69.42 | 94.13 | 53.93 | 77.53 | 52.52 | 78.90 | 65.71 |
| Seedance 1.5 [seedance] | 0.20 | 7.69 | 49.18 | 0.375 | 12.28 | 0.602 | 0.342 | 39.23 | 97.31 | 56.60 | 94.81 | 54.45 | 74.88 | 44.21 | 75.61 | 59.91 |
| Vidu Q3 [vidu] | 0.24 | 8.47 | 42.75 | 0.360 | 12.65 | 0.586 | 0.351 | 39.05 | 99.43 | 71.28 | 95.01 | 55.38 | 76.73 | 40.90 | 79.57 | 60.24 |
| HappyHorse 1.0 [happyhorse] | 0.18 | 5.62 | 58.29 | 0.420 | 12.48 | 0.564 | 0.352 | 42.45 | 99.62 | 74.17 | 95.02 | 55.56 | 77.00 | 50.37 | 80.27 | 65.32 |
5.1 Experimental Setup
We evaluate 20 representative video world models: 6 camera-driven, 7 action-driven, and 7 language-driven models. For each case, we construct the model-facing input in the model’s native format using the interface adaptation described in section˜3 and evaluate the generated video with the protocols in section˜4. All 20 models are evaluated on the static-scene track. The dynamic-interaction track requires observable third-person subject-scene interaction and therefore includes WorldPlay [worldplay], LingBot-World [lingbot], and all seven language-driven models. The other five action-driven models either do not support third-person subject control or cannot control a visible third-person subject reliably, and are therefore excluded from the dynamic-interaction track. Camera-driven models are excluded because their interfaces control only the camera, and Goal Completion is evaluated only on language-driven models. To avoid compromising model performance, we use each model’s default resolution, video length, and other inference settings whenever applicable. For streaming models with flexible generation lengths, we constrain the output to 100–200 frames to prevent quality drift in excessively long generations from biasing the evaluation. Some closed-source commercial language-driven systems apply proprietary prompt enhancement before video generation. Consistent with the functional formulation in section˜3, we retain such default preprocessing as part of the native end-to-end pipeline. Detailed inference settings and task eligibility are provided in section˜8; unsupported tasks are marked with dashes in the result tables.
5.2 Static-Scene Track Results
Table˜2 reports Camera Control and Scene Revisit together with the task-agnostic general metrics. The task scores diagnose Control Adherence and Spatial Consistency, whereas the general metrics characterize Visual Quality; reporting them separately reveals that these capabilities do not necessarily improve together.
Camera Control.
Among camera-driven models, the strongest results come from methods that reconstruct 3D priors and reproject them to target views: NeoVerse [neoverse] scores 97.33 and InSpatio-World [inspatio] scores 85.94. ReCamMaster [recammaster] and FantasyWorld [fantasyworld], which encode camera poses as learned tokens or embeddings, obtain much lower Camera Control scores of 38.64 and 18.46 despite competitive General averages of 80.97 and 80.23. WorldPlay [worldplay] is the strongest action-driven model at 92.74. Language-driven models are less precise, with Hailuo 2.3 [hailuo] achieving the strongest score of 63.29, consistent with the difficulty of expressing ordered, complex viewpoint changes through natural-language instructions.
Scene Revisit.
NeoVerse and InSpatio-World both achieve 1.000 Revisit Success and lead the camera-driven group with Scene Revisit scores of 89.25 and 85.90, respectively. Among action-driven models, WorldPlay performs best with 0.790 Revisit Success and a score of 72.51, while Hailuo 2.3 leads the language-driven group with 0.505 and 48.70. The remaining gap reflects failures either to return to the initial viewpoint or to recover its geometry, appearance, and content after the round trip.
5.3 Effect of the Input Translation Multiplier
We evaluate NeoVerse [neoverse] under the same camera controls while varying only the translation multiplier of its input trajectory. As the multiplier increases from to , the Camera Control score decreases from 98.25 to 95.32, the Scene Revisit score from 90.98 to 88.37, and the General average from 80.42 to 75.05; all five general metrics decline, with Photometric Consistency falling from 80.17 to 62.45. The results confirm that larger image-space displacements make both controlled generation and scene preservation more difficult. This ablation therefore motivates the pre-generation image-space displacement alignment in section˜4.2, which is applied to all reported camera-driven results.
| Translation Multiplier | Camera Control | Scene Revisit | 3D Cons. | Photo. Cons. | Temp. Flick. | Aesth. Quality | Imag. Quality | Average | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T. Err. | R. Err. | Score | Success | PSNR | LPIPS | SSIM | Score | Task | General | Overall | ||||||
| 0.10 | 0.002 | 0.43 | 98.25 | 1.000 | 22.68 | 0.128 | 0.704 | 90.98 | 99.91 | 80.17 | 95.41 | 54.06 | 72.53 | 94.62 | 80.42 | 87.52 |
| 0.50 | 0.003 | 0.51 | 97.84 | 1.000 | 22.17 | 0.134 | 0.683 | 90.11 | 99.54 | 76.33 | 94.21 | 53.39 | 72.47 | 93.98 | 79.19 | 86.59 |
| 0.75 | 0.004 | 0.56 | 97.57 | 1.000 | 22.00 | 0.137 | 0.676 | 89.80 | 99.07 | 73.86 | 93.81 | 53.03 | 72.30 | 93.69 | 78.41 | 86.05 |
| 1.00 | 0.005 | 0.60 | 97.33 | 1.000 | 21.72 | 0.141 | 0.662 | 89.25 | 98.19 | 70.83 | 93.53 | 52.72 | 72.19 | 93.29 | 77.49 | 85.39 |
| 2.00 | 0.015 | 0.91 | 95.32 | 1.000 | 21.32 | 0.151 | 0.641 | 88.37 | 96.32 | 62.45 | 92.93 | 52.00 | 71.54 | 91.85 | 75.05 | 83.45 |
5.4 Dynamic-Interaction Track Results
| Model | Subject Control | Terrain Inter. | Object Inter. | Social Inter. | Physical Reaction | Goal Completion | Subject Cons. | Motion Smooth. | Aesth. Quality | Imag. Quality | Average | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Task | General | Overall | |||||||||||
| Action-driven | |||||||||||||
| WorldPlay [worldplay] | 49.75 | 27.49 | 33.75 | 51.40 | 26.91 | – | 88.06 | 98.09 | 54.19 | 68.76 | 37.86 | 77.28 | 57.57 |
| LingBot-World [lingbot] | 55.47 | 24.33 | 25.94 | 60.37 | 33.43 | – | 94.98 | 98.89 | 60.86 | 71.69 | 39.91 | 81.61 | 60.76 |
| Language-driven | |||||||||||||
| Kling 2.5 [kling] | 28.40 | 35.95 | 27.70 | 66.80 | 31.99 | 48.25 | 96.00 | 99.48 | 56.86 | 71.83 | 39.85 | 81.04 | 60.45 |
| Veo 3.1 [veo] | 37.28 | 44.71 | 75.96 | 85.10 | 61.76 | 85.30 | 92.07 | 99.12 | 58.22 | 72.65 | 65.02 | 80.52 | 72.77 |
| Hailuo 2.3 [hailuo] | 36.49 | 61.57 | 67.01 | 72.45 | 63.84 | 78.86 | 93.25 | 99.31 | 57.74 | 72.47 | 63.37 | 80.69 | 72.03 |
| Wan 2.6 I2V [wan] | 29.02 | 49.21 | 44.56 | 66.40 | 48.15 | 76.39 | 94.46 | 98.15 | 56.30 | 74.72 | 52.29 | 80.91 | 66.60 |
| Seedance 1.5 [seedance] | 32.51 | 53.83 | 37.91 | 72.09 | 47.60 | 76.24 | 92.14 | 98.87 | 56.72 | 70.84 | 53.36 | 79.64 | 66.50 |
| Vidu Q3 [vidu] | 27.67 | 64.39 | 71.59 | 81.91 | 61.23 | 78.26 | 92.78 | 98.76 | 57.24 | 73.25 | 64.18 | 80.51 | 72.35 |
| HappyHorse 1.0 [happyhorse] | 33.11 | 56.30 | 65.70 | 76.17 | 47.01 | 85.33 | 92.69 | 98.85 | 57.37 | 73.23 | 60.60 | 80.54 | 70.57 |
Table˜4 reveals whether a model follows subject-motion control and whether it can react correctly.
Subject Control.
The direct action interfaces provide more precise subject control: LingBot-World [lingbot] scores 55.47 and WorldPlay [worldplay] scores 49.75, compared with the best language-driven score of 37.28 from Veo 3.1 [veo]. Even the action-driven results remain far from saturated, with failures often converting the requested subject motion into camera motion or leaving the scene static.
Terrain Interaction.
Vidu Q3 [vidu] and Hailuo 2.3 [hailuo] lead with 64.39 and 61.57, whereas the best action-driven score is 27.49. For action-driven models, the large drop from Subject Control to Terrain Interaction shows that horizontal control adherence does not guarantee vertical terrain adaptation.
Object Interaction.
Veo 3.1 and Vidu Q3 lead with 75.96 and 71.59, while the best action-driven score is 33.75. Common failures leave the contacted object unchanged or allow the subject to pass through it.
Social Interaction.
Veo 3.1 achieves the highest score of 85.10, followed by Vidu Q3 at 81.91; the best action-driven score is 60.37. Failures typically leave nearby agents unresponsive or allow the controlled subject to move through them without avoidance or yielding.
Physical Reaction.
Hailuo 2.3 leads with 63.84, followed by Veo 3.1 and Vidu Q3 at 61.76 and 61.23; the best action-driven score is 33.43. Action-driven generations often execute subject control while leaving unstable or contacted objects unchanged, exposing the gap between explicit control and inherent reactivity.
Goal Completion.
HappyHorse 1.0 [happyhorse] and Veo 3.1 achieve the strongest results at 85.33 and 85.30. Kling 2.5 [kling] scores only 48.25 despite having the highest General average among language-driven models, showing that visual quality does not guarantee grounded goal-directed execution.
5.5 Cross-Task Diagnostic Analysis
Figure˜8 consolidates the capability split across the three interfaces. Camera-driven models provide the strongest camera control and scene revisiting but do not support dynamic interaction. Action-driven models control designated subjects more precisely, yet this advantage does not consistently transfer to the scene-conditioned reactions induced by those controls. Language-driven models perform better on interaction and goal-directed tasks but follow composed camera and subject controls less faithfully. No model combines broad coverage with consistently strong performance, leaving current interfaces complementary but incomplete.
The split is also obscured by general visual metrics. For example, the language-driven General averages occupy a narrow range of 79.64–81.04, while their Task averages range from 39.85 to 65.02. Likewise, ReCamMaster and FantasyWorld retain strong General averages despite weak Camera Control scores. These gaps show that the four diagnostic levels capture distinct capabilities. In particular, strong Visual Quality or Control Adherence does not guarantee World Reactivity, motivating separate reporting of the four levels.
5.6 Human Alignment of Checklist Evaluation
| Task | Checklist Items | Spearman | PLCC |
|---|---|---|---|
| Goal Completion | 1,303 | 0.8960 | 0.9017 |
| Physical Reaction | 1,425 | 0.8838 | 0.8750 |
| Object Interaction | 1,527 | 0.8251 | 0.8541 |
| Social Interaction | 1,538 | 0.7019 | 0.7103 |
| Overall | 5,793 | 0.8614 | 0.8583 |
We validate the VLM judge on the four tasks evaluated using checklists. The validation set contains 800 evaluation instances and 5,793 checklist items, with 200 instances per task. Three human annotators independently label each item from the same 10 temporally ordered frames shown to the VLM, and the majority vote defines the binary reference label. For each instance, the human and VLM scores are computed as the respective fractions of satisfied checklist items. Table˜5 reports Spearman’s and PLCC for each task separately and across all 800 evaluation instances combined. Across all 800 instances, Spearman’s is 0.8614 and PLCC is 0.8583, showing strong agreement between the VLM judge and human evaluation.
5.7 Backend Stability with DA3 Reconstruction
To assess sensitivity to the reconstruction backend, we rerun the geometry-based metrics with Depth Anything 3 (DA3) [da3] while keeping the benchmark inputs and model outputs fixed. Tables˜6 and 7 report the corresponding results. Scores for the four tasks evaluated using checklists remain unchanged.
On the static-scene track, the mean absolute relative change in Overall score is 3.09% across all 20 models: 0.44% for camera-driven, 3.08% for action-driven, and 5.36% for language-driven models. Most variation is concentrated in Camera Control, while Scene Revisit and the general metrics change little. NeoVerse, WorldPlay, and Hailuo 2.3 remain the leading models in their respective groups; the camera- and action-driven rankings are fully preserved, with only closely matched language-driven models exchanging positions.
On the dynamic-interaction track, DA3 affects only Subject Control, Terrain Interaction, and their geometry-dependent aggregates. The mean absolute relative change in Overall score is 0.57%, the maximum change is 1.16%, and all within-paradigm rankings are preserved. Together, these small changes and stable rankings show that the main model comparisons do not depend on a particular reconstruction backend.
| Model | Camera Control | Scene Revisit | 3D Cons. | Photo. Cons. | Temp. Flick. | Aesth. Quality | Imag. Quality | Average | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| T. Err. | R. Err. | Score | Success | PSNR | LPIPS | SSIM | Score | Task | General | Overall | ||||||
| Camera-driven | ||||||||||||||||
| TrajectoryCrafter [trajectorycrafter] | 0.13 | 1.10 | 81.97 | 1.000 | 19.06 | 0.240 | 0.544 | 82.99 | 95.63 | 62.83 | 92.56 | 52.18 | 67.03 | 82.48 | 74.05 | 78.27 |
| ReCamMaster [recammaster] | 0.47 | 4.01 | 38.19 | 0.815 | 16.03 | 0.352 | 0.415 | 68.04 | 99.18 | 81.97 | 95.51 | 54.58 | 73.45 | 53.12 | 80.94 | 67.03 |
| Voyager [voyager] | 0.30 | 4.04 | 60.03 | 0.995 | 16.89 | 0.386 | 0.467 | 76.33 | 89.81 | 27.44 | 93.04 | 53.59 | 63.69 | 68.18 | 65.51 | 66.85 |
| FantasyWorld [fantasyworld] | 1.05 | 7.76 | 18.24 | 0.555 | 15.39 | 0.323 | 0.373 | 55.51 | 98.79 | 75.60 | 95.90 | 57.11 | 73.98 | 36.87 | 80.28 | 58.58 |
| NeoVerse [neoverse] | 0.01 | 0.72 | 96.93 | 1.000 | 21.70 | 0.141 | 0.661 | 89.22 | 97.53 | 70.83 | 93.53 | 52.72 | 72.19 | 93.07 | 77.36 | 85.22 |
| InSpatio-World (1.3B) [inspatio] | 0.08 | 1.21 | 86.91 | 1.000 | 20.77 | 0.225 | 0.605 | 85.84 | 97.79 | 64.88 | 93.68 | 53.77 | 73.60 | 86.38 | 76.74 | 81.56 |
| Action-driven | ||||||||||||||||
| Hunyuan-GameCraft [gamecraft] | 0.13 | 11.92 | 39.17 | 0.375 | 13.37 | 0.542 | 0.369 | 41.26 | 94.00 | 53.93 | 93.91 | 54.98 | 71.70 | 40.21 | 73.70 | 56.96 |
| Astra [astra] | 0.71 | 7.97 | 15.97 | 0.360 | 12.18 | 0.600 | 0.309 | 37.89 | 97.09 | 78.42 | 96.37 | 51.71 | 71.78 | 26.93 | 79.07 | 53.00 |
| WorldPlay [worldplay] | 0.08 | 1.06 | 88.31 | 0.780 | 18.69 | 0.269 | 0.544 | 72.52 | 98.55 | 81.06 | 95.90 | 53.92 | 73.31 | 80.41 | 80.55 | 80.48 |
| Yume 1.5 [yume15] | 0.32 | 3.14 | 59.69 | 0.270 | 12.37 | 0.614 | 0.359 | 33.40 | 97.74 | 67.53 | 95.18 | 53.42 | 75.27 | 46.55 | 77.83 | 62.19 |
| LingBot-World [lingbot] | 0.19 | 6.28 | 54.43 | 0.620 | 15.63 | 0.340 | 0.409 | 59.17 | 99.42 | 81.52 | 96.53 | 59.87 | 75.01 | 56.80 | 82.47 | 69.64 |
| Infinite-World [infiniteworld] | 0.25 | 2.38 | 64.49 | 0.655 | 14.30 | 0.389 | 0.345 | 57.76 | 99.87 | 92.09 | 95.99 | 56.27 | 76.40 | 61.12 | 84.12 | 72.62 |
| Matrix-Game 3.0 [matrixgame3] | 0.17 | 4.27 | 68.93 | 0.860 | 13.83 | 0.486 | 0.372 | 64.23 | 95.61 | 30.14 | 93.39 | 48.80 | 73.39 | 66.58 | 68.27 | 67.43 |
| Language-driven | ||||||||||||||||
| Kling 2.5 [kling] | 0.31 | 6.93 | 43.37 | 0.315 | 13.50 | 0.535 | 0.380 | 38.13 | 99.78 | 87.87 | 97.57 | 56.28 | 75.70 | 40.75 | 83.44 | 62.10 |
| Veo 3.1 [veo] | 0.89 | 8.20 | 21.27 | 0.447 | 12.30 | 0.595 | 0.345 | 43.03 | 98.73 | 73.61 | 95.42 | 55.49 | 76.57 | 32.15 | 79.96 | 56.06 |
| Hailuo 2.3 [hailuo] | 0.28 | 5.18 | 54.61 | 0.510 | 13.56 | 0.518 | 0.388 | 49.00 | 98.94 | 79.88 | 95.32 | 56.30 | 75.52 | 51.81 | 81.19 | 66.50 |
| Wan 2.6 I2V [wan] | 0.38 | 6.12 | 46.50 | 0.485 | 13.29 | 0.527 | 0.355 | 46.88 | 99.27 | 69.43 | 94.13 | 53.93 | 77.53 | 46.69 | 78.86 | 62.78 |
| Seedance 1.5 [seedance] | 0.78 | 7.78 | 28.47 | 0.370 | 12.29 | 0.601 | 0.343 | 39.01 | 95.96 | 56.60 | 94.82 | 54.45 | 74.89 | 33.74 | 75.34 | 54.54 |
| Vidu Q3 [vidu] | 0.49 | 8.48 | 30.69 | 0.360 | 12.67 | 0.582 | 0.352 | 39.14 | 99.13 | 71.27 | 95.01 | 55.38 | 76.73 | 34.92 | 79.50 | 57.21 |
| HappyHorse 1.0 [happyhorse] | 0.40 | 5.55 | 45.66 | 0.415 | 12.45 | 0.564 | 0.351 | 42.16 | 99.52 | 74.17 | 95.02 | 55.57 | 76.99 | 43.91 | 80.25 | 62.08 |
| Model | Subject Control | Terrain Inter. | Object Inter. | Social Inter. | Physical Reaction | Goal Completion | Subject Cons. | Motion Smooth. | Aesth. Quality | Imag. Quality | Average | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Task | General | Overall | |||||||||||
| Action-driven | |||||||||||||
| WorldPlay [worldplay] | 47.51 | 23.05 | 33.75 | 51.40 | 26.91 | – | 88.06 | 98.09 | 54.19 | 68.76 | 36.52 | 77.28 | 56.90 |
| LingBot-World [lingbot] | 53.43 | 21.63 | 25.94 | 60.37 | 33.43 | – | 94.98 | 98.89 | 60.87 | 71.70 | 38.96 | 81.61 | 60.29 |
| Language-driven | |||||||||||||
| Kling 2.5 [kling] | 26.78 | 37.50 | 27.70 | 66.80 | 31.99 | 48.25 | 96.00 | 99.49 | 56.86 | 71.83 | 39.84 | 81.05 | 60.45 |
| Veo 3.1 [veo] | 35.12 | 42.19 | 75.96 | 85.10 | 61.76 | 85.30 | 92.07 | 99.12 | 58.21 | 72.65 | 64.24 | 80.51 | 72.38 |
| Hailuo 2.3 [hailuo] | 35.16 | 59.34 | 67.01 | 72.45 | 63.84 | 78.86 | 93.26 | 99.31 | 57.74 | 72.47 | 62.78 | 80.70 | 71.74 |
| Wan 2.6 I2V [wan] | 26.95 | 48.55 | 44.56 | 66.40 | 48.15 | 76.39 | 94.46 | 98.15 | 56.30 | 74.71 | 51.83 | 80.91 | 66.37 |
| Seedance 1.5 [seedance] | 29.41 | 48.28 | 37.91 | 72.09 | 47.60 | 76.24 | 92.14 | 98.87 | 56.71 | 70.84 | 51.92 | 79.64 | 65.78 |
| Vidu Q3 [vidu] | 26.15 | 62.56 | 71.59 | 81.91 | 61.23 | 78.26 | 92.78 | 98.76 | 57.24 | 73.25 | 63.62 | 80.51 | 72.07 |
| HappyHorse 1.0 [happyhorse] | 31.62 | 53.81 | 65.70 | 76.17 | 47.01 | 85.33 | 92.69 | 98.86 | 57.37 | 73.23 | 59.94 | 80.54 | 70.24 |
6 Conclusion
We presented WorldExam, a unified hierarchical benchmark for diagnosing video world models beyond visual quality and explicit instruction fulfillment. It distinguishes direct fulfillment of explicitly specified controls or targets from scene-conditioned reactions and detailed goal-directed execution that must be inferred from the initial scene. This distinction is instantiated through four diagnostic levels, eight tasks, and 1,474 test cases. Interface adaptation presents shared cases in the native formats of camera-, action-, and language-driven models, while the static-scene and dynamic-interaction tracks restrict evaluation to compatible interfaces rather than treating unsupported tasks as failures.
Evaluation of 20 representative models reveals a clear capability split. Camera-driven models provide the most precise camera control and scene revisiting; action-driven models control subjects more precisely but often leave terrain, objects, nearby agents, and physical processes unresponsive; and language-driven models perform better on interaction and goal-directed tasks but follow composed controls less faithfully. No evaluated model combines broad task coverage with consistently strong performance, showing that high visual quality and explicit instruction fulfillment do not guarantee inherent reactivity. Strong agreement between human and VLM checklist scores, together with stable model rankings under an alternative reconstruction backend, supports the reliability of these findings.
The current scope is bounded by the capabilities of available model interfaces. Dynamic-interaction evaluation requires reliable third-person subject control, and Goal Completion remains limited to language-driven models. Moreover, the metrics assess observable end-to-end behavior rather than determining where reasoning occurs or establishing that the video generator itself has learned an internal causal representation. For closed-source commercial systems, proprietary prompt enhancement may contribute to scene grounding and execution planning. Future extensions to broader interfaces, longer-horizon interactions, and more intervention-based settings would provide stronger tests of persistent world understanding. Within its current scope, WorldExam identifies where controllable video models succeed, where their generated worlds remain unresponsive, and which capabilities must be developed jointly.
7 Gallery
8 Per-Model Inference Settings
To preserve each model’s native performance, we use its default resolution, video length, and other inference settings whenever applicable. For streaming models with flexible generation lengths, we constrain the output to 100–200 frames to prevent quality drift in excessively long generations from biasing the evaluation. Table˜8 reports the resulting resolution and frame count for each model.
Eligibility for the dynamic-interaction track requires reliable control of a visible third-person subject. Among action-driven models, only WorldPlay [worldplay] and LingBot-World [lingbot] satisfy this requirement; the other five either do not support third-person subject control or cannot provide it reliably. All evaluated language-driven models accept third-person subject-motion prompts, whereas camera-driven interfaces control only the camera. Goal Completion is language-only within the dynamic-interaction track.
| Model | Backend | Resolution | Frames | Dynamic Track |
| Camera-driven | ||||
| TrajectoryCrafter [trajectorycrafter] | Local | 49 | – | |
| ReCamMaster [recammaster] | Local | 81 | – | |
| Voyager [voyager] | Local | 49 | – | |
| FantasyWorld [fantasyworld] | Local | 81 | – | |
| NeoVerse [neoverse] | Local | 81 | – | |
| InSpatio-World (1.3B) [inspatio] | Local | 81 | – | |
| Action-driven | ||||
| Hunyuan-GameCraft [gamecraft] | Local | 132 | ✗ TPV | |
| Astra [astra] | Local | 161 | ✗ FPV only | |
| WorldPlay [worldplay] | Local | 125 | ✓ TPV | |
| Yume 1.5 [yume15] | Local | 145 | ✗ FPV only | |
| LingBot-World [lingbot] | Local | 161 | ✓ TPV | |
| Infinite-World [infiniteworld] | Local | 161 | ✗ FPV only | |
| Matrix-Game 3.0 [matrixgame3] | Local | 177 | ✗ FPV only | |
| Language-driven | ||||
| Kling 2.5 [kling] | API | 153 | ✓ TPV prompt | |
| Veo 3.1 [veo] | API | 192 | ✓ TPV prompt | |
| Hailuo 2.3 [hailuo] | API | 141 | ✓ TPV prompt | |
| Wan 2.6 I2V [wan] | API | 150 | ✓ TPV prompt | |
| Seedance 1.5 [seedance] | API | 97 | ✓ TPV prompt | |
| Vidu Q3 [vidu] | API | 121 | ✓ TPV prompt | |
| HappyHorse 1.0 [happyhorse] | API | 123 | ✓ TPV prompt | |
9 Qualitative Examples of the Eight Tasks
The examples below instantiate all eight tasks in a shared visual format. The highlighted image is the initial frame, followed by four temporally ordered generated frames. The bottom panel shows the shared control intent or high-level goal. For tasks evaluated using checklists, the checklist is shown only to explain the evaluation protocol and is withheld from the input.
9.1 Camera Control
Camera Control measures whether an ordered composition of camera motions follows the prescribed directions and temporal order without unintended drift. In this case, the camera tilts down, pans left, and moves left.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx11.png&exp=1785888000&sig=63e11fa5c457919f4a62fd2ea6bcd1ad16166df073ce871b1c14ee8a62b95704)
9.2 Subject Control
Subject Control applies atomic controls to a designated third-person subject. Here, the subject should move forward, right, and left in order while remaining visually identifiable.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx12.png&exp=1785888000&sig=e11de5d3762446dc0a49e64dc94e59e0736a3367929e84d532e28c08b5e0ff59)
9.3 Scene Revisit
Scene Revisit couples a round-trip camera motion with a spatial-memory requirement: the camera should return to the initial viewpoint while preserving the scene.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx13.png&exp=1785888000&sig=558ae17e3fb47927c2c8f63d2a5c88c03023ded825a66a63704278d0a80e4625)
9.4 Terrain Interaction
Terrain Interaction specifies only horizontal subject motion. The model must infer the vertical adaptation required by the visible terrain—in this case, traversing the stairs while continuing forward.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx14.png&exp=1785888000&sig=decd69ed71967d7fbc019150f26a4b6539983360707302f7c4f4376fc9b486bd)
9.5 Object Interaction
Object Interaction tests whether contact with a designated target causes a response consistent with the target’s physical type. Here, the worker pushes a bus cart forward into a lightweight sign.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx15.png&exp=1785888000&sig=2647b737c221d464fe909c4cb4a9a6abb9675326a39fe19ddd149ba2ba4c6ddd)
9.6 Social Interaction
Social Interaction evaluates whether nearby agents respond plausibly when a controlled subject enters their social or safety space. Here, a sedan enters a storefront crossing with two pedestrians and a shopping cart.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx16.png&exp=1785888000&sig=e1f5a60032e5a913aaf49275d044ca585019ca05c87518e6a3e8f0a2e140b0b1)
9.7 Physical Reaction
Physical Reaction evaluates the temporal development of a physical process, not merely whether contact occurs. Here, the woman remains stationary while a towel-loaded laundry basket, whose center of mass overhangs the washer edge, begins to tip and fall.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx17.png&exp=1785888000&sig=5ab289402b6416ee545bc0a3522d6a7c6c8dd347904d19b9b26ec532cc2b5988)
9.8 Goal Completion
Goal Completion provides a high-level goal instead of an atomic control sequence. The model must ground the goal in the initial scene and generate coherent execution steps toward the desired target.
![[Uncaptioned image]](/api/img-proxy?u=https%3A%2F%2Farxiv.org%2Fhtml%2F2608.02603v1%2Fx18.png&exp=1785888000&sig=81726d74a915afb02558f411f071225dad576f9f2407e35d7fe5ea7c6737ccc1)