HappyWorld-Bench is a benchmark for evaluating the simulated worlds that AI generates — so-called world models. It asks a single question: not whether a world looks plausible at a glance, but whether it stays coherent when you step inside it, walk around, touch things, and change them. What's new is that it pulls together three strands that have been evaluated in isolation until now — video, spatial, and embodied (robotic) — under a single yardstick, and it looks at the reliability of interaction rather than raw generation quality.
A world model is an internal simulator that an agent (an AI that decides and acts on its own) uses to imagine, in its head, "if I act like this, how will the environment change?" The appeal is obvious: you can picture the future and pick good actions without actually going through trial and error. But for that imagining to be useful, one premise has to hold. The world must stay consistent as the camera moves, respond plausibly when you reach in and touch it, and preserve its state even after you've wandered for a long time or come back to a place you'd already seen. In short, prediction has to survive interaction.
Think of a movie set. In a single camera shot it can look like a flawless mansion, but open the door and walk in, and the back might be plywood. What we want is a real house — one you can actually move into, where moving a piece of furniture leaves a gap, and where the furniture you moved is still where you left it when you return to the room.
The trouble is that, until now, the ruler for measuring this "real house" has been fragmented. One benchmark looked only at controllability, another only at physical plausibility, another only at memory or long-horizon consistency. On top of that, video, spatial, and embodied models take different inputs and pursue different goals, making them hard to compare on the same scale. So the question — "does this go beyond surface-level quality to become a world you can actually trust?" — had no single screen that could answer it.
HappyWorld-Bench is structured to answer three questions in order: what to measure (the W1-W6 capability hierarchy), how to measure it (three tracks), and by what yardstick (a two-pronged evaluation scheme). The three interlock to stitch fragmented evaluation into one.
First, what — the W1-W6 capability hierarchy. It defines world modeling in six levels, with each higher level stacking more demands on the ones below.
| Level | Name | Core question |
|---|---|---|
| W1 | Perceptual world | Does it render a world with correct semantics and spatial structure from the input (including short-range continuity)? |
| W2 | Interactive world | When an action is applied, does the state change in line with local physics and causality? |
| W3 | Persistent world | After long traversal, occlusion, and revisiting, are spatial and object identities preserved? |
| W4 | Programmable world | Does only the intended target get modified while the rest stays untouched? |
| W5 | Scalable world | Can it maintain and coordinate an infinitely expanding world shared by multiple agents? |
| W6 | Universal world | Does it integrate all of the above into one, fully reproducing and generalizing reality? |
The key is that it demands more than just making good images (W1). It requires responding plausibly to outside intervention (W2), remembering that state over time (W3), and even modifying only a specific part on purpose (W4) — higher-order reliability. On top of that sit multi-agent scaling (W5) and universal integration (W6), so it reads as a difficulty ladder.
Second, how — three independent tracks. They share the same W1-W6 taxonomy, but each measures different aspects to fit the model's form. The overall scale is as follows.
| Track | What it examines | Scale |
|---|---|---|
| Video world model | Are the observations after the input spatially and temporally consistent? | 1,138 prompts (data covers W1-W5) |
| Spatial world model | Does the exported scene support valid physical manipulation? | 300 scenes (266 for evaluation + 34 for expansion, W1-W5) |
| Embodied world model | From a first-person view, how do the robot, scene, and objects change in response to actions? | 254 tests (W2-W4) |
Combined, the three tracks come to 1,692 instances, and 14 video, 9 spatial, and 8 embodied models were evaluated. The data spans nine domains — robotics, urban, indoor, nature, traffic, gaming, everyday life, industry, and fantasy — with interaction lengths ranging from brief clips to 60-second sequences, and first-frame resolutions from under 1MP to over 8MP.
Third, by what yardstick — a two-pronged scheme of human judgment and automatic metrics. All three tracks run HappyWorld-Arena to collect human A/B comparisons (placing two outputs side by side and choosing which is better), then convert these into within-track Elo scores. Much like a chess ranking, blind head-to-head records summarize overall human preference. On top of that, automatic metrics that capture behavioral accuracy are added to finely measure specific capabilities. Each track fleshes out the W hierarchy to suit its own model characteristics — temporal consistency for video, physical manipulability for spatial, predicting the outcome of an embodied agent's actions, and so on.
The figures the paper reports in its abstract and body are laid out here track by track, in the same frame (headline result → exposed limitation). Because each track uses a different scale, we do not sum them or compare them against one another. In particular, the video track's Elo scores and the spatial track's accuracy (%) have different units entirely, so there's no way to bundle them into a "which track is better" claim.
Video track. The two leading systems posted Arena Elo scores of 1263 and 1206 respectively, showing a relative edge in human preference. But in the capability-by-capability evaluation, even the strongest system revealed limits in keeping the world state consistent and responding to actions and interventions as intended. Consistency slipped especially in long rollouts (the process where the AI keeps extending its own next-scene predictions over a long stretch) and in revisit situations.
Spatial track. Existing spatial world models fell short on physical plausibility, editability, and scene expansion. Even the top result reached only 70.14% placement accuracy and 73.33% edit execution. In other words, even when they produced seemingly complete scenes, there was a substantial gap in actually manipulable geometry and in reflecting the intended edits.
Embodied track. The models struggled to preserve state across multi-step actions and failed to respond accurately to changed action conditions or physics rules. It's a weakness of the same grain as video and spatial: the outcome of a prior action doesn't carry over properly into the next scene.
What all three tracks point to together is clear. Generation fidelity has already reached an impressive level, but the ability to reliably hold onto state through repeated change is still unfinished across all three strands. A world that looks good and a world you can trust turned out to be different problems.
The limitations the paper states are as follows.
The insight of this paper applies not only to AI agents, but to any "state management" work that deals with complex information. Picture, for instance, managing a project with many stakeholders tangled together. As requests, decisions, dependencies, and schedules pass back and forth over many turns, the states fixed earlier must be preserved without wobbling. That is exactly what the AI project managers ONTLE assigns to each client project do: they maintain the state of the world that is the project.
The scene most worth stress-testing is the "decision change." When a client reverses a decision, does the PM change only that one thing accurately and — without touching the unrelated existing decisions — keep them consistent through the conversation that follows? For example, while applying the change "let's push the deployment schedule back a week," an unrelated earlier decision like "the design mockup is locked to option A" must not quietly get shaken loose. This touches directly on what the paper examines in W4 (modify only the intended target, leave the rest) and W3 (preserve state even on revisit). If the response looks smooth but, a few steps later, something you'd agreed not to change has changed, that is precisely a world that collapses on revisit.
The verification approach is worth borrowing too. The two-pronged structure — preference evaluation where humans blind-compare two outputs, plus automatic metrics that pinpoint a specific capability — maps directly onto QA for ONTLE's own deliverables. In short, this paper shows how to measure "looks pretty" and "can be trusted" as separate things. It's a distinction worth engraving for every team putting AI agents into a product.
It's a simulator in which an agent internally predicts "if I act like this, how will the environment change?" That prediction has to hold up consistently through interaction for it to be useful in actual planning and action selection.
They aren't scores but a hierarchy (ladder) of capabilities. Demands accumulate from perception (W1) through interaction (W2), persistence (W3), programmability (W4), and scalability (W5) up to universal integration (W6). The measured scope differs by track: video covers W1-W4 by published results (W1-W5 in the data), spatial covers W1-W5, and embodied covers W2-W4.
No. It's a diagnostic tool that exposes the reliability gaps. It points to the direction that models should be evaluated on state consistency and the accuracy of their responses to actions and interventions — not on fidelity — while closing the gaps is left to follow-up research.