WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity
Yuxue Yang, Shuyao Shang, Jiahe Wang, Zitong Zhou, Liang Tan, Junhan Zeng, Ruizhi Li, Junyan Li, Yu Liu, Xiao Yang, Yong Li, Jun Zhu, Hongsheng Li, Tieniu Tan, Lue Fan, Zhaoxiang Zhang
cs.CV
2026-08-04
WorldExam tests 20 video world models on 1,474 cases of inherent reactivity; strong visuals and instruction-following don't mean the generated world actually reacts.
Controllable video generation models are increasingly sold as world models: give one an initial frame plus a camera, action, or language instruction and it should predict a plausible future. But today's benchmarks stop at two things: how good the video looks, and whether the model rendered the action or interaction outcome the prompt explicitly asked for. What separates a world model from a clip generator is the reaction the input never spells out: a subject climbing stairs should rise and fall with the terrain, an obstacle should produce contact or avoidance, another agent should yield when its space is invaded. These are scene-conditioned consequences, not direct depictions of the instruction. WorldExam, led by CASIA with CUHK and Tsinghua, targets exactly this blind spot.
The benchmark is a four-level diagnostic ladder: Visual Quality, Control Adherence, Spatial Consistency, and World Reactivity. The first three stay in explicit territory: perceptual quality, whether the camera or subject obeys, whether the world still holds together when the camera returns to a previous viewpoint. Only the fourth level probes implicit reaction.
Cross-paradigm comparison rests on two steps. Every controllable behavior is decomposed into atomic control units (W/S/A/D and camera primitives), then interface-adapted: the same control intent is rendered as an SE(3) trajectory for camera-driven models, as discrete actions for action-driven models, and as natural language for language-driven models. The key design choice lives in World Reactivity: the model-facing input carries only the explicit control or high-level goal, leaving every scene-conditioned reaction unstated and forcing the model to infer it. This level splits into five tasks: Terrain Interaction, Object Interaction, Social Interaction, Physical Reaction (gravity, collision, friction, fluid), and Goal Completion (language-only, goal but no steps).
The full benchmark is 1,474 cases across eight tasks, organized in two tracks. The static-scene track (Camera Control, Scene Revisit) runs on all three paradigms; the dynamic-interaction track (Subject Control plus the five reactivity tasks) runs only on action- and language-driven models that can reliably control a third-person subject; Goal Completion is language-only. Scoring is split too: tasks with geometric ground truth are lifted to 3D with VGGT-Ω and evaluated in the reconstructed space; the four semantic tasks go to a GPT-5.5 VLM judge that scores against per-case checklists. Camera-driven models also get a pre-generation image-space displacement calibration so they are not penalized for translating more pixels than peers.
Twenty models (6 camera, 7 action, 7 language) were tested. Headline: strong visuals do not imply reactivity.
On the static track, Camera Control scores split widely. NeoVerse and InSpatio-World, which reconstruct 3D priors and reproject, score 97.33 and 85.94; ReCamMaster and FantasyWorld, which encode camera pose as learned tokens, sit at 38.64 and 18.46 despite General averages above 80. WorldPlay leads action-driven models at 92.74; Hailuo 2.3 tops language-driven at just 63.29. Scene Revisit: NeoVerse 89.25, InSpatio-World 85.90. Across the seven language-driven models the General average is bunched in 79.64–81.04, while Task averages range 39.85–65.02, so visual metrics cannot separate them.
On the dynamic track, reactivity is the filter. Action-driven models control subjects more precisely (LingBot-World 55.47, WorldPlay 49.75, both above the best language score of 37.28 from Veo 3.1) but collapse on reaction: best Physical Reaction score only 33.43. Veo 3.1 leads Social Interaction at 85.10 and Object Interaction at 75.96; Hailuo 2.3 takes Physical Reaction at 63.84; Vidu Q3 leads Terrain Interaction at 64.39; HappyHorse 1.0 and Veo 3.1 tie on Goal Completion around 85.3. The sharpest contrast is Kling 2.5: highest General average among language models, yet only 48.25 on Goal Completion.
Supporting checks hold up. Pushing NeoVerse's translation multiplier from 0.10× to 2.00× drops Camera Control from 98.25 to 95.32 and Photometric Consistency from 80.17 to 62.45. The VLM judge agrees with human labels at Spearman ρ=0.8614 and PLCC=0.8583 across 800 instances, weakest on Social Interaction at ρ=0.7019. Swapping VGGT-Ω for DA3 changes Overall by a mean 3.09% on the static track and 0.57% on the dynamic track, with rankings preserved.
Three signals for practitioners. First, you cannot pick a video world model off VBench-style visual scores: the language-driven group is nearly indistinguishable on General metrics and only reactivity tasks separate them. Second, every paradigm pays a different cost: camera-driven models cannot interact, action-driven models control precisely but leave the world inert, language-driven models act out interactions but flub composed control. No model covers the board. Third, teams deploying these models for games, robotics, or embodied AI should expect that reaction is often cosmetic: terrain, contact, social conflict, and physics still regularly freeze or break.
The authors concede several bounds: the dynamic-interaction track requires reliable third-person subject control, excluding most action-driven and all camera-driven models; Goal Completion is language-only; metrics capture end-to-end observable behavior and cannot prove the model learned an internal causal representation; and proprietary prompt enhancement in closed systems may inflate scene grounding. A few more doubts: the VLM judge carries its own bias (Social Interaction human alignment only ρ=0.70), absolute scores on semantic tasks may not be stable, the 100–200 frame horizon leaves long-horizon physics untested, and interface adaptation may systematically undersell action-driven models whose discrete actions map awkwardly onto SE(3).