World Model Arena: Interactive Evaluation Benchmark

• World Model Arena is an online platform that evaluates interactive video world models using standardized action protocols and diverse test cases.
• It unifies assessments of visual quality, control alignment, and world consistency through side-by-side battles and live leaderboard comparisons.
• The benchmark addresses fairness by standardizing inputs and metrics, enabling reproducible insights into both perceptual and functional model capabilities.

World Model Arena denotes, in the narrow sense, the online comparison platform introduced alongside WorldMark at warena.ai, where interactive video world models are matched in side-by-side battles under standardized conditions, and, in a broader research sense, an emerging benchmark paradigm in which world models are evaluated as controllable simulators rather than as isolated video generators (Xu et al., 23 Apr 2026). The concept arises from a convergence of several lines of work: standardized action-conditioned video benchmarking, embodied-functional evaluation, long-horizon stability analysis, and arena-style human or proxy preference comparison. Across these efforts, the central premise is that a world model must be assessed not only for visual plausibility, but also for action obedience, persistence, physical coherence, planning utility, and reproducibility under shared test conditions.

1. Historical emergence and scope

The immediate origin of the proper name World Model Arena is the WorldMark benchmark suite, which couples an offline standardized benchmark with a public online platform for head-to-head comparison of leading interactive world models (Xu et al., 23 Apr 2026). In that formulation, the arena is not a separate theory of world modeling; it is the public operational layer that exposes benchmarked systems through side-by-side battles and a live leaderboard. WorldMark explicitly distinguishes the two roles: WorldMark is the benchmark, while World Model Arena is the online system that makes the benchmark visible and continuously comparable.

At the same time, multiple contemporaneous projects extend the same evaluative logic beyond a single platform. WorldArena frames world-model evaluation around both perceptual and functional embodied utility (Shang et al., 9 Feb 2026). WorldArena 2.0 expands that frame along modality, functionality, and platform, including visuotactile prediction and sim-to-real evaluation (Shang et al., 18 May 2026). WR-Arena evaluates world models as internal simulators for semantic action following, long-horizon forecasting, and planning (Team et al., 26 Mar 2026). WorldRoamBench focuses on long-horizon stability under continuous interactive roaming, emphasizing action, vision, physics, and memory (Xu et al., 30 Jun 2026). This suggests that “World Model Arena” has become a useful umbrella label for a family of benchmark designs whose shared objective is to make world-model progress measurable under interactive, comparable, and diagnostically rich conditions.

Benchmark or platform	Primary emphasis	Distinctive feature
WorldMark / World Model Arena	Comparable evaluation of interactive Image-to-Video world models	Unified action-mapping layer and online side-by-side battles
WorldArena	Embodied world models	Joint perceptual and functional evaluation
WorldArena 2.0	Embodied world models across modality, functionality, and platform	Visuotactile, RL-environment, and sim-to-real extension
WR-Arena	Next world simulation	Action simulation, long-horizon forecast, and planning
WorldRoamBench	Long-horizon open-world interaction	Explicit action, vision, physics, and memory decomposition

2. The fairness problem that made arena-style benchmarking necessary

The principal motivation for WorldMark is a fairness problem: interactive video generation models such as Genie, YUME, HY-World, Matrix-Game, HY-GameCraft, and Open-Oasis had been evaluated on private scenes, private trajectories, heterogeneous control interfaces, and often non-comparable metrics (Xu et al., 23 Apr 2026). Under those conditions, a high score can reflect an easier scene, a more forgiving action sequence, or a control format tailored to the model. WorldMark’s core intervention is therefore not merely to add more metrics, but to create a common playing field in which models receive the same scenes, the same action sequences, the same semantic instructions, and the same evaluation dimensions.

The technical mechanism that makes this possible is the unified action interface. WorldMark defines a canonical six-primitive vocabulary—W, S, A, D, L, and R—with duration parameterization, and translates that shared action language into each model’s native interface through per-model adapters (Xu et al., 23 Apr 2026). This preserves semantic equivalence while respecting architectural heterogeneity: YUME uses natural-language captions, HY-World and HY-GameCraft use structured 6-DoF pose-related inputs, Genie uses gamepad-style controls, Matrix-Game uses action-function calls, and Open-Oasis uses a 25-dimensional action vector. The adapters additionally calibrate movement and yaw rates so that a command such as forward motion for a specified duration corresponds as closely as possible to the same intended motion across systems.

The same standardization logic appears in other arena-oriented benchmarks, albeit with different task definitions. WorldArena argues that prior embodied world-model evaluation suffered from a perception-only bias, largely treating world models as video generators rather than as data engines, policy evaluators, or action planners (Shang et al., 9 Feb 2026). WR-Arena similarly argues that short-horizon next-state prediction and visual fidelity are insufficient proxies for simulation quality when the target use case is reasoning and decision support (Team et al., 26 Mar 2026). WorldRoamBench makes an analogous point for long-horizon interaction: trajectory-level action metrics alone are inadequate because they hide per-frame failures and confound memory with control error (Xu et al., 30 Jun 2026). Across all of these works, the common methodological claim is that evaluation becomes meaningful only when benchmark conditions are normalized and failure modes are disentangled.

3. Benchmark construction and online arena mechanics

WorldMark’s offline benchmark is a hierarchical test suite of 500 cases derived from 50 reference images, expanded into 100 test images by including paired first-person and synthesized third-person views, and combined with 15 standardized action sequences (Xu et al., 23 Apr 2026). The scene distribution spans first-person and third-person viewpoints, photorealistic and stylized scenes, and category types including Nature, City, and Indoor. The stylized subset includes oil painting, Ukiyo-e, cyberpunk, and Minecraft-style imagery, which reduces bias toward purely photographic inputs.

The action suite is stratified into three difficulty tiers. Easy cases are single-segment 20 s trajectories; Medium cases are two-segment 40 s trajectories; Hard cases are three-segment 60 s trajectories, including patrol routes or full 360° panoramic rotations (Xu et al., 23 Apr 2026). This hierarchy separates immediate compliance from sustained long-horizon behavior. A model may follow a short forward command while still failing under extended patrol or repeated rotation. WorldMark also uses a VLM-based filtering step so that selected trajectories remain physically plausible for the scene, for example by rejecting lateral movement when the image contains walls or other obstacles that make such motion implausible.

The online World Model Arena inherits these standardized inputs and exposes them as side-by-side battles and a live leaderboard (Xu et al., 23 Apr 2026). The battle format is important because it reveals qualitative differences that aggregate metrics may obscure, including visual fidelity, action obedience, temporal coherence, and scene persistence. The live leaderboard makes the benchmark dynamic rather than static: new models and new evaluations can be incorporated into a public ranking as the field evolves.

An analogous live-evaluation logic appears in VisionArena, although that system targets vision-LLMs rather than world models. VisionArena is built on the Chatbot Arena paradigm, collects 230K real user conversations, includes 30K anonymous pairwise battles, and uses a Bradley-Terry model to infer model strength from preference data (Chou et al., 2024). It also distills arena logs into VisionArena-Bench, a 500-prompt offline benchmark intended to approximate the live leaderboard. This provides a methodological template for arena-style evaluation: live human or user-mediated battles, anonymous pairwise comparison, a public ranking mechanism, and an offline distilled benchmark that tracks the online arena. This suggests that World Model Arena is part of a broader move toward hybrid online-offline evaluation infrastructures for interactive AI systems.

4. Evaluation dimensions: from visual quality to functionality, reasoning, physics, and memory

WorldMark organizes evaluation into three complementary dimensions: Visual Quality, Control Alignment, and World Consistency (Xu et al., 23 Apr 2026). Visual Quality uses the LAION aesthetic predictor for Aesthetic Quality and MUSIQ for Imaging Quality. Control Alignment reconstructs camera motion from generated video with DROID-SLAM and compares estimated trajectories with intended trajectories using scale-invariant translation error and geodesic rotation error. World Consistency combines DROID-SLAM reprojection error with VLM-based State Consistency, Content Consistency, and Style Consistency. The intent is to prevent a model from scoring well through a single narrow strength such as frame attractiveness while failing in long-horizon control or scene stability.

WorldArena generalizes this logic for embodied world models through 16 metrics across 6 sub-dimensions: visual quality, motion quality, content consistency, physics adherence, 3D accuracy, and controllability (Shang et al., 9 Feb 2026). It supplements those perceptual axes with three functional settings—world models as data engines, policy evaluators, and action planners—and adds human subjective evaluation through scalar ratings and pairwise win-rate comparison. The benchmark further defines EWMScore, a composite metric formed by linearly normalizing all 16 metrics to $[0,100]$ and taking their arithmetic mean. A central result is the perception-functionality gap: high visual quality does not necessarily imply strong embodied-task capability.

WorldArena 2.0 extends the same structure in three directions. Along the modality axis, it adds a visuotactile pipeline built on UniVTAC, with tactile prediction evaluated by PSNR and SSIM and downstream manipulation evaluated by task success rate (Shang et al., 18 May 2026). Along the functionality axis, it evaluates world models as interactive RL environments rather than only as predictors or planners. Along the platform axis, it extends beyond simulation-only testing to RoboTwin 2.0, LIBERO, and the AgileX Split-Type ALOHA real robot. Its conclusion is that simulation metrics alone are not a reliable proxy for real-world usability, especially for content consistency and controllability.

WR-Arena reorients the evaluative vocabulary toward Action Simulation Fidelity, Long-horizon Forecast, and Simulative Reasoning and Planning (Team et al., 26 Mar 2026). Action Simulation Fidelity asks whether the model follows semantically meaningful, multi-step instructions in agent-centered and environment-centered settings. Long-horizon Forecast uses Transition Smoothness and Generation Consistency, including the explicitly defined Multi-round Smoothness (MRS) metric and an additive penalty for long-run degradation. Simulative Reasoning and Planning evaluates whether the model improves goal-directed action selection in structured tasks from Language Table and open-ended tasks from Agibot. A main empirical claim is that only semantically actionable rollouts improve planning reliably.

WorldRoamBench makes the decomposition more explicit for long interactive sessions by evaluating Action, Vision, Physics, and Memory (Xu et al., 30 Jun 2026). Its Action score combines strict per-frame action accuracy, partial action accuracy, and a trajectory-level TrajScore. Its Vision score combines average aesthetic and imaging quality with segment-based drift metrics designed to catch non-monotonic mid-rollout collapse. Its Physics dimension is controllability-gated and evaluates mechanics, optics, and 3D consistency. Its Memory protocol separates scene memory from subject memory and uses transition-localized 3D point-cloud reconstruction plus SAM2 and Qwen3-VL-PLUS-based subject assessment. The benchmark’s stated result is that, across 600+ test cases and 10+ open/closed-source models, no model reliably satisfies all four dimensions.

5. Systems, architectures, and environment substrates represented in the arena landscape

The systems that populate this evaluative landscape are architecturally heterogeneous. WorldMark itself targets six representative interactive video models—YUME 1.5, Matrix-Game 2.0, HY-World 1.5, HY-GameCraft, Open-Oasis, and Genie 3—precisely because they expose different native control interfaces and therefore require semantic standardization before comparison (Xu et al., 23 Apr 2026). WorldRoamBench broadens the model roster further to include systems such as Happy Oyster, LingBot-World, Lyra 2.0, SANA-WM, Matrix-Game 3.0, Yume 1.5, and minWM, and reports that no single model dominates action, visual stability, physics, and memory simultaneously (Xu et al., 30 Jun 2026).

Several recent systems illustrate the architectural directions that arena benchmarks are trying to discriminate. MultiGen introduces an explicit external memory that persists independently of the model’s context window and decomposes generation into Memory, Observation, and Dynamics modules (Po et al., 3 Mar 2026). Its state is formalized as $S_t = (M, p_t, o_{t-L+1:t})$, where $M$ is an editable geometric map and $p_t$ is player pose, and it supports real-time multiplayer rollouts by letting multiple per-player modules read from and write to shared memory. This is directly relevant to arena evaluation because it operationalizes editability, reproducibility, and shared-world consistency rather than relying on implicit frame-history memory alone.

DreamForge-World 0.1 Preview represents a different design point: a low-compute, consumer-GPU, real-time interactive preview system adapted from the LongLive 1 autoregressive video stack with a residual action pathway inspired by Matrix-Game (Ayupov et al., 29 Jun 2026). It supports live keyboard and mouse control, multimodal initialization, mid-stream reprompting, dual-view operation, and minute-scale continuation at native 480p, reaching 14–15 FPS on a single RTX 4090 with the LightTAEW 2.1 decoder path. The paper is explicit that it is not yet a memory-complete or frontier-quality world simulator. In arena terms, it is a pragmatic baseline for real-time controllability under limited compute rather than a benchmark leader on persistence.

HY-World 2.0 pushes in the opposite direction toward unified 3D world creation and reconstruction (HY-World et al., 15 Apr 2026). It accepts text, single-view images, multi-view images, and videos; performs panorama generation with HY-Pano 2.0, trajectory planning with WorldNav, world expansion with WorldStereo 2.0, and world composition with WorldMirror 2.0; and outputs explorable 3D Gaussian Splatting scenes plus meshes. Its runtime layer, WorldLens, provides collision detection, engine-agnostic rendering, automatic IBL lighting, and interactive exploration. This positions the system closer to an offline 3D world-construction stack than to a purely autoregressive video simulator, but it remains relevant to arena evaluation because it targets reconstruction fidelity, navigability, and interaction-ready world assets.

Beyond video-native systems, several works propose alternative world substrates that are likely to matter for future arena design. Web World Models define a hybrid state split $S_t = (S_t^\phi, S_t^\psi)$, where deterministic web code implements the Physics Layer and an LLM implements the Imagination Layer (Feng et al., 29 Dec 2025). The design uses typed web interfaces, deterministic generation for object permanence, and graceful degradation. World Craft formalizes a playable world as $\mathcal{G} = (M, A, L, P)$ and uses World Scaffold and World Guild to convert text into executable AI Town-like environments (Sun et al., 14 Jan 2026). Arena 4.0 generates human-centric navigation environments from text prompts or 2D floorplans through Arena-gen, a semantic 3D model database, and a full ROS 2 benchmarking stack (Shcherbyna1 et al., 2024). These systems suggest that future “World Model Arena” instances may compare not only video rollouts, but also persistent, typed, editable, and simulator-integrated worlds.

6. Conceptual tensions, misconceptions, and open problems

A recurring misconception in the literature is that world-model quality can be inferred from video realism alone. WorldArena rejects this directly by showing a perception-functionality gap, where high visual quality does not necessarily translate into strong embodied-task capability (Shang et al., 9 Feb 2026). WR-Arena reaches an analogous conclusion for planning: visually plausible rollouts do not reliably help action selection unless they are semantically actionable and causally informative (Team et al., 26 Mar 2026). WorldRoamBench makes the same point in long-horizon interaction, noting that high visual quality does not imply good controllability, and that stricter physics can even lower action scores when collisions block unrealistic motion commands (Xu et al., 30 Jun 2026).

Another tension concerns the relation between control fidelity and physical plausibility. WorldMark’s standardized action mapping is designed to equalize motion intent across heterogeneous interfaces, but that does not remove deeper differences in how models internalize action and environment dynamics (Xu et al., 23 Apr 2026). WorldRoamBench explicitly documents the tradeoff: a model that enforces collision strongly may score worse on unconstrained path execution even though its behavior is more physically grounded (Xu et al., 30 Jun 2026). This suggests that arena evaluation must distinguish instruction obedience from simulation validity, rather than treating them as interchangeable.

Memory is a further unresolved issue. MultiGen argues that long-horizon reproducibility, editability, and multiplayer coherence all fail when state is only implicit in a finite visual history, and therefore treats explicit external memory as a missing primitive for world models (Po et al., 3 Mar 2026). DreamForge-World, by contrast, is explicit that persistent spatial memory is currently missing, so revisited scenes may be re-synthesized rather than preserved (Ayupov et al., 29 Jun 2026). WorldRoamBench formalizes this distinction at the benchmark level by decoupling memory evaluation from action error through point-cloud-based revisit analysis (Xu et al., 30 Jun 2026). This suggests that future arena leaderboards may need to report memory-aware and control-aware scores separately rather than collapsing them into a single scalar.

The broader trajectory of the field points toward increasingly multi-axis evaluation. WorldArena 2.0 adds tactile sensing, RL-environment use, and sim-to-real transfer because vision-only, offline, simulator-only benchmarks no longer match the ambitions of embodied world models (Shang et al., 18 May 2026). HY-World 2.0, Web World Models, and World Craft indicate that world models are also becoming more diverse in representational form: 3DGS worlds, code-defined persistent web worlds, and executable structured scenes are all candidates for future comparison (HY-World et al., 15 Apr 2026). A plausible implication is that the next generation of World Model Arena platforms will need to compare models across multiple substrates—video rollouts, explicit world state, embodied utility, and environment editability—while preserving the central requirement established by WorldMark: identical conditions, semantically equivalent control, and evaluation protocols that reveal rather than hide failure modes.