World Model Evaluation

We introduce reference-free measures for evaluating the physical consistency of generated videos, combining relative and absolute approaches to assess fidelity. Although tools like WorldGym or WorldEval enable robotic simulation via video generation, physical fidelity gaps often prevent these environments from accurately reproducing real-world task success rates of VLA models. Unlike existing evaluation methods, which require costly human voting (Elo) or unavailable ground-truth references (FVD), our approach utilizes DROID-SLAM and SEA-RAFT to quantify physical inconsistencies, motivated by WorldScore. Videos filtered using our relative consistency assessment show an improvement in task success rates of over 8%, effectively narrowing the simulation-to-reality gap. Furthermore, our absolute assessment enables spatio-temporal localization, providing visualization of when and where physical artifacts occur.

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

Scaling robot policy learning for autonomous surgery is challenging, as expert demonstrations are expensive and in vivo exploration poses substantial safety risks. Surgical world models address this by generating realistic, action-conditioned future frames from an initial observation, but existing methods exhibit two persistent failure modes: spatial interaction incoherence, where visible instrument contact fails to induce spatially consistent tissue deformation, and temporal fidelity collapse, where prediction errors compound across autoregressive rollouts and progressively corrupt visual quality. We present SurgVista, a surgical world model that mitigates both failures through two training recipes. Deformation Consistency Regularization extracts scene-point trajectories from training videos and enforces cross-frame coherence through latent contrastive learning, strengthening physically consistent instrument-tissue dynamics. Drift Adaptation Training mitigates long-horizon drift by perturbing conditioning frames with online prediction residuals and photometric augmentations calibrated to long-horizon drift statistics, sustaining visual fidelity over extended rollouts. To enable rigorous evaluation, we further introduce SurgWorld-Bench, featuring diverse procedure types, long-range rollouts, and decoupled metrics for instrument-motion accuracy and tissue-response fidelity. Extensive experiments show that SurgVista consistently outperforms state-of-the-art methods across visual quality, temporal consistency, and interaction fidelity, with gains widening as the prediction horizon grows.

Evaluating generalist robot manipulation policies in the real world is expensive, slow, and difficult to scale. Action-conditioned video world models offer a scalable alternative by simulating policy rollouts. Autoregressive rollouts accumulate compounding errors, observations across multiple camera views must remain mutually consistent, and the evaluator must generalize to policies whose behaviors lie outside the training distribution. We address these challenges with SC3-Eval, a self-consistent video generation recipe that adapts a pre-trained video foundation model into an accurate policy evaluator by enforcing three complementary forms of consistency. First, forward-inverse dynamics consistency jointly trains the model to predict frames from actions and to recover actions from frames, anchoring generated rollouts to a physically plausible action manifold and counteracting the drift a forward-only model cannot penalize. Second, cross-view consistency trains the model to inpaint each camera view from the other, keeping the multi-camera observation coherent over long rollouts without any explicit memory mechanism. Third, test-time consistency reuses the inverse dynamics mode at inference as a per-action-chunk uncertainty signal that terminates rollouts whose generated frames drift away from the requested actions. We also demonstrate SC3-Eval rollouts reproduce the failure modes that policies exhibit in real-world rollouts, supporting fine-grained diagnostic comparison rather than aggregate ranking alone. Across seven real-world vision-language-action policies, SC3-Eval attains a closed-loop Pearson correlation of $0.929$ and MMRV of $0.119$, outperforming three strong prior video-model-based baselines, and generalizes to new tasks.

World models are widely used in robotic and agentic engineering control systems due to their ability to learn latent dynamics for planning and decision-making. As these systems are increasingly deployed in safety-critical settings, understanding their robustness under adversarial conditions has become essential. However, existing evaluations lack a unified benchmark for testing adversarial threats across the policy, value, and latent-dynamics levels of world-model agents. To fill this gap, we present ARB4WM, a unified evaluation framework for pre-deployment robustness and risk assessment of world-model agents under visual perturbations. ARB4WM defines five white-box loss objectives across these three levels and studies their effects when combined with single-step or multi-step perturbation strategies and temporal attack modes, including full-frame, half-sequence, and sparse-frame exposure. Specifically, we evaluate four Dreamer-style agents across 20 tasks from MetaWorld and the DeepMind Control Suite under different loss objectives, perturbation strategies, and temporal attack modes. Results show that attacks targeting value estimation, latent representations, and RSSM dynamics can be as damaging as direct policy disruption, and that early or frequent perturbations are especially harmful, while input-level defenses provide limited recovery under adaptive attacks. These findings suggest that safety, risk, and reliability assessment for world models should cover multiple component-oriented attack objectives and temporal exposure protocols rather than relying solely on action-space robustness. Source code is available at https://github.com/zaoanguai/ARB4WM.

World-model synthesis aims to turn interaction experience into an internal model of environment dynamics. Existing symbolic approaches often fit observed transitions or mixtures of local rules, but they do not produce a complete executable program that can run independently of the real environment. We present Mind-Studio, a framework that synthesizes executable pygame-style world models from state-action-next-state trajectories using large language models. Mind-Studio combines entropy-selected traces with a lightweight game skill file containing object, action, and static scene information extracted from screenshots. We evaluate synthesis quality with a K-step lookahead fidelity protocol that compares generated world-model rollouts against Real-ALE rollouts from the same state. On Montezuma's Revenge, Mind-Studio improves chosen-action next-state prediction from 0.3% for PoE-World to 48.7% while verifying 5 of 8 subgoals; across Alien, Assault, and Skiing, it achieves stronger branch-level fidelity than prior learned lookahead sources.

World models have become a central abstraction in modern AI. The term now refers to several different objects: action-conditioned environment models, latent imagination models, future-video predictors, interactive neural simulators, latent predictive representations, and synthetic-data engines. Evaluation has broadened along with the term. Recent papers measure video realism, perceptual similarity, instruction following, physical plausibility, policy ranking, executability, planning success, and downstream policy improvement. This produces both metric diversity and a recurring problem of claim/evidence mismatch: papers sometimes make a stronger claim about what their model is useful for than their evaluation can establish. This paper surveys the recent literature and argues that, for models presented as world models for embodied decision-making, the more decisive issue is not whether the model generates visually convincing videos, but whether it supports reliable interventional reasoning, policy evaluation, planning, and policy optimization under intervention, policy-induced distribution shift, and long-horizon rollout. We organize the survey using an L0--L7 ladder spanning visual plausibility to policy optimization utility, noting that the levels cut across several orthogonal axes and so form an evidential hierarchy rather than a single scalar. The framework foregrounds interventional action fidelity, closed-loop rollout validity, reward/value prediction, policy-ranking agreement, optimization lift, model exploitability, and uncertainty calibration, with a minimal feasible reporting set for real-robot settings.

Reactive capability is a key property of data-driven behavior world model simulators for autonomous driving simulation systems. With this capability, simulated world agents can respond feasibly to autonomous vehicle (AV) behaviors that differ from the log. However, existing behavior simulation benchmarks do not directly measure reactive capability. They often let the simulator jointly control the AV and surrounding agents and evaluate realism through log similarity or open-loop prediction metrics. In this work, we introduce ReactSim-Bench for evaluating the reactive capability of behavior world model simulation in autonomous driving. We decouple the control of agents and the AV, using AV behaviors that differ from the log and require agents to respond as independent AV inputs. To obtain these AV behaviors, we construct a pipeline that uses an AV planner model to generate candidate behaviors and filters the data using rules and manual verification. Collision metrics, map-based metrics, and kinematic feasibility metrics are used to evaluate the safety and rule compliance of reactive responses. We construct 2,636 test scenarios with three categories and conduct a systematic evaluation of state-of-the-art models across multiple architectures, including Transformer-based, diffusion-based, and next-token-prediction-based models. We further analyze how replan frequency affects performance and provide insights for future studies.

The potential impacts of world models (WMs, i.e., learned simulators) on robotics are far-reaching -- policy evaluation, policy improvement, and test-time planning -- all with limited real-world interaction. To unlock these downstream capabilities, a WM needs to jointly satisfy three desiderata: $\textit{(i)}$ fidelity (i.e., producing simulated trajectories that correlate with reality), $\textit{(ii)}$ consistency (i.e., producing simulated trajectories that are coherent over long horizons), and $\textit{(iii)}$ efficiency (i.e., producing simulated trajectories quickly). We propose WEAVER (World Estimation Across Views for Embodied Reasoning): a WM architecture that simultaneously achieves all three desiderata, providing state-of-the-art results on robotic manipulation tasks. WEAVER is a multi-view WM trained to predict future latents and reward values via a flow-matching loss. We distill the key design decisions across model architecture, memory, and prediction objectives required to unlock the kinds of long-horizon dynamic manipulation tasks that have confounded prior world modeling approaches. We apply WEAVER in robotic hardware, demonstrating its effectiveness at policy evaluation ($ρ$=0.870 correlation with real-world success rate), policy improvement (real-world success rate improvement of $38\%$ on top of the $π_{0.5}$ robot foundation model), and test-time planning (real-world success rate improvement of $14\%$ with a $5-10\times$ speedup over prior WMs). WEAVER also demonstrates better performance than prior WMs when evaluated on out-of-distribution scenarios. Code, models, and videos at: https://arnavkj1995.github.io/WEAVER/ .

Scale buys interpolation; structure buys certifiable transfer. A world model's average error does not say whether a particular rollout can be trusted, or for how long. For equivariant latent world models we give a predictability certificate: a computable region spanning configuration, horizon, and resolution. Under exact equivariance, rollout error is invariant over the monoid generated by k primitive symmetries and is certified from the k generators (Theorem A); universal orbit-flatness over equivariant targets characterizes equivariance at the function level (Lemma 2), so an unconstrained architecture cannot certify the property by construction. Approximate orbit-transfer defects propagate by the finite-time Lyapunov spectrum (Theorem B): expanding channels give a logarithmic horizon $T_j(ε)\sim\log(1/ε)/λ_j$, neutral channels accumulate recurrent defect linearly, and contracting channels accumulate a bounded nonzero floor. Exact conserved charge values are certified to all horizons only at zero defect; with one-step defect $η$, charge-value error grows at most as $Tη$. Empirically, on a 40-dimensional learned model a $\mathbb{Z}_N$-equivariant network recovers the full Lyapunov spectrum ($R^2=0.98$-$0.99$) where dense and recurrent baselines fail. A cone/adapted-metric certificate reads an a-priori horizon off the model's own Jacobian, tight on uniformly hyperbolic dynamics and self-abstaining elsewhere; the resulting horizon improves a budgeted re-observation decision. For public non-equivariant world models the tangent spectrum gives a training-free candidate horizon, paired with a held-out divergence cross-check that abstains or corrects when the learned loop over-promises.

We introduce WorldOlympiad, a benchmark for diagnosing video-based world models across physical faithfulness, geometric consistency, and interaction fidelity. While existing benchmarks often focus on visual quality, semantic alignment, or short-term temporal coherence, they provide limited insight into whether generated videos obey physical rules, preserve coherent 3D structure, and sustain controllable interactions over long horizons. To address this gap, WorldOlympiad decomposes world-model evaluation into three complementary dimensions. The physical track uses object segmentation and MLLM-as-judge to assess whether generated videos follow interpretable rules in mechanics, thermal phenomena, and material properties. The geometry track reconstructs generated videos with Gaussian splatting and evaluates structural consistency, cross-view coherence, and camera-trajectory alignment. The interaction track assesses whether generated rollouts follow complex action prompts and maintain smooth, coherent transitions across consecutive video chunks. WorldOlympiad further covers three major downstream scenarios, including gaming, robotics, and general real-world videos, capturing diverse challenges from interactive control and embodied manipulation to open-domain motion and camera dynamics. Together, these tracks and scenarios form a scalable and interpretable evaluation suite that exposes failure modes beyond generic video quality. Experiments on state-of-the-art models reveal substantial gaps in physical reasoning, 3D consistency, and long-horizon interaction, underscoring the need for more structured evaluation protocols for generative world models.

We present \textbf{Echo-Memory}, a controlled study of memory mechanisms in action-conditioned world models. These models generate multi-segment videos from a first frame, text prompt, and camera-action sequence, but their central failure is often memory rather than local image synthesis: after the camera leaves and returns, the scene or salient object may silently change. Existing memory designs are hard to compare because gains are entangled with backbone, training, retrieval, and evaluation differences. Echo-Memory fixes the action-to-video interface and varies only how history is stored and read by the generator. Under a shared video diffusion backbone, optimizer, camera-action representation, sampler, and evaluation pipeline, we compare raw context, compression-based memory, spatial summaries with different read-out paths, and state-space recurrence. This matched matrix separates four otherwise conflated axes: \emph{capacity}, \emph{compression}, \emph{read-out}, and \emph{recurrence}. We also evaluate memory through a three-branch protocol: replay quality, in-domain loop revisit, and open-domain return probes. The branches routinely disagree, showing that replay fidelity is not a sufficient proxy for remembering a world. Three findings follow. Raw context is a strong capacity baseline and improves open-domain return far more than it improves replay metrics. Compactness is not a free substitute for capacity: aggressive spatial and hybrid-compression memories lose the salient evidence needed for return. Finally, block-wise state-space recurrence is the strongest open-domain return mechanism in our matrix, showing that the structure of implicit memory matters as much as the decision to use it. These results provide a compact protocol for studying memory in action world models beyond isolated replay metrics.

Large language model (LLM)-based agents are increasingly used in interactive textual environments, from web navigation and code editing to tool use and long-horizon dialogue. Yet many remain largely reactive, mapping observations to actions without an explicit model of how these environments are structured and evolve. This motivates text world models (TWMs): transition models over textual states that, given a state and a candidate action, predict the resulting webpage, terminal output, API response, or user reply, thereby supporting planning, efficient learning, and principled evaluation. We systematically review text world models for LLM-based agents, organized around a formal framework and the agent lifecycle: (1) Foundations, defining text world models and characterizing them by state representation and grounding domain; (2) Construction, taxonomizing LLM-as-WM and code-as-WM paradigms and reviewing methods for building them; (3) Application, examining how world models support agents at training time through experience synthesis and at inference time through planning, verification, and adaptation; and (4) Evaluation, covering both evaluation of the world model itself and its use as an evaluation environment for agents. We aim to consolidate this rapidly developing area, clarify its design space, and highlight open challenges for future research.

Latent world models can achieve accurate latent prediction yet still differ substantially in downstream planning performance. We argue that a key source of this discrepancy lies in the structure of action-induced latent transitions. We formalize action-identifiability through Bayes inverse risk, characterizing how much uncertainty about an action remains after observing the transition it induces. Model-predicted transitions can become highly self-decodable while encoding a domain-specific action relationship that fails to transfer to real environment transitions. We characterize this mismatch through cross-domain inverse transfer, instantiated as the Action-Consistency Transfer Matrix (ATM), a $2\times2$ inverse-risk matrix over real and predicted transition domains. Across TwoRoom, PushT, and OGBench-Cube, true-transition action-identifiability tracks downstream planning substantially better than standard prediction loss, while the full ATM reveals highly self-decodable yet cross-domain-inconsistent predicted transitions. Controlled interventions on the two domains further produce distinct transition structures and planning outcomes, supporting this decomposition. The same diagnostics also support lightweight model screening, reaching 98.81% pairwise ranking accuracy for candidates separated by more than 5% success.

Controllable video world models target interactive world exploration, where models must faithfully execute explicit action commands while preserving visual quality and temporal coherence. However, most existing approaches rely on continuous camera trajectories as action conditions, which often lead to unreliable action following, especially under complex motion sequences. In this work, we identify action representation entanglement as a key bottleneck in controllable video generation, and show that continuous camera representations lead to high feature similarity across distinct motion patterns, degrading action controllability. Based on this insight, we propose DisCo, a controllable video world model that conditions generation on a compact set of discrete action primitives to improve action separability. We further introduce DisCoBench, a comprehensive benchmark for evaluating the ability of models in short-term, long-horizon, and highly dynamic exploration scenarios. Extensive experiments demonstrate that DisCo achieves significantly more reliable action following while preserving visual quality.

Vision-language-action (VLA) policies operate in a closed loop in real-world robot tasks: a robot observes the scene, executes an action chunk, and conditions its next decision on the resulting observation. However, most existing world models for robot action evaluation are limited to open-loop prediction along pre-collected action trajectories. This prevents them from supporting closed-loop VLA evaluation, where each action chunk must be conditioned on the observation generated by the previous execution. To address this gap, we propose PiL-World, a chunk-wise world model designed for policy-in-the-loop VLA evaluation. Given the current observation and the action trajectory rolled out by a VLA policy, PiL-World generates multi-view future observations that are consistent with the VLA rollout and match the image inputs required by the policy. By alternating between VLA inference and world-model prediction, PiL-World enables closed-loop evaluation without real robot execution at every step. To improve rollout fidelity, PiL-World conditions video generation on action-derived visual control from head-view robot motion and latent histories that encode task execution context, while jointly predicting complementary multi-view observations. Beyond successful teleoperated demonstrations, it also learns from failed execution trajectories, helping the imagined rollouts better match the distribution of real policy executions. We evaluate PiL-World on three real dual-arm manipulation tasks. PiL-World generates imagined rollouts that are highly consistent with real robot executions. More importantly, compared with the baseline, it reduces the error between VLA success rates measured in real-world rollouts and those estimated through closed-loop world-model evaluation from 63.2% to 12.0%.

World models, learned generative models that predict how an environment evolves, have become a promising tool for sample-efficient robot learning. Yet how robust they are to environmental variability remains poorly understood. To address this, we conduct a systematic study using vision-based quadrotor navigation as a testbed problem, training DreamerV3-based world models under varying levels of environmental randomness and evaluating them across all levels through cross-environment validation, spanning both Self-Supervised Learning (SSL) pretraining and Reinforcement Learning (RL) fine-tuning. We then deploy all world models and associated navigation policies on a real quadrotor in unseen environments, including an open-loop run where the model receives just 2.5s of real sensory input before all sensors are cut off, leaving the system to navigate entirely in imagination over a 12m traverse. Our results show that world model robustness during SSL pretraining is a strong predictor of sim-to-real transfer: every model that generalized well in cross-environment SSL validation deployed successfully in the real world, passing through gaps as narrow as 0.67m, whereas the model that dominated simulation policy evaluation failed on the real platform. We further identify (a) the discrete latent size and (b) the training-sequence length as the dominant factors governing world model quality.

We present OSCAR, a precise action-conditioned video world model that generalizes across different robot embodiments and enables robot policy evaluation. Existing video world models face three main challenges for real-world robot evaluation: limited scenario diversity in current robot training datasets, imprecise action following, and poor generalization across embodiments for broad adoption. We tackle these challenges from two perspectives. At its core is a large-scale standardized data pipeline that curates, filters, and deduplicates broad robotics and egocentric human datasets, yielding a clean joint-training dataset that spans diverse tasks, scenarios, actions, and robot embodiments. To condition the video model, we adopt 2D kinematic skeleton rendering as a unified conditioning representation that generalizes across different robot arms or even human hands. We finetune the Cosmos-Predict2.5-2B model on a single GH200 GPU. Our model achieves significant improvement on action following, appearance quality, and motion consistency, compared to existing baselines, which either have a much larger model size or require more GPUs. We further deploy OSCAR to evaluate robot policies from RoboArena. Extensive experiments demonstrate the significant correlation between our virtual policy evaluation in OSCAR and real-world evaluation, paving the way for the future where robot policies can be purely evaluated in virtual generated worlds.

Video world models are increasingly used in robotic manipulation, yet existing benchmarks mostly evaluate them under valid, feasible, and safe instructions. We introduce RoboTrustBench, a benchmark for evaluating the trustworthiness of video world models under four scenarios: Normal, Constraint-Sensitive, Counterfactual, and Adversarial. Built from real-world DROID episodes, RoboTrustBench contains 1,207 expert-validated instruction-image pairs and a six-dimensional evaluation protocol with 13 fine-grained criteria. Evaluating seven representative video world models with human and MLLM assessment, we find that current models often generate visually coherent videos, but struggle with constraint reasoning, counterfactual grounding, physical interaction, and unsafe-instruction suppression. These results show that visual quality and surface-level instruction following are insufficient for trustworthy robotic video world modeling.

With rapid development of large language models and diffusion-based content generation, world modeling has attracted increasing research attention, benefiting various downstream domains such as game engines, embodied AI, autonomous driving, etc. Through explicitly incorporating user actions into world state transition, recent literature empowers world modeling with interactivity in an action-conditioned video or 3D generation paradigm, further enhancing controllability over world evolutions and facilitating users to freely traverse, manipulate, navigate, and personalize the state evolution. In this paper, we aim to systematically review recent research trends, technical developments, evaluation benchmarks, and also propose future potential directions in interactive world modeling. Specifically, we first summarize recent efforts and trends in terms of application scenarios, world state evolution, and scene modality. Afterwards, we delve into three crucial technical challenges, including action-conditioned controllability, long-horizon interactions and memory, and action-following responsiveness for real-time interactivity. Furthermore, we also thoroughly compare existing benchmarks and metrics in four specific application fields: open-world exploration, game engine, autonomous driving, and robotics. Finally, we discuss several promising future directions in achieving next-generation interactive world modeling. The corresponding repository is publicly available at: https://github.com/liujiuming123/Awesome-Interactive-World-Model.

Recent advancements in video-based world models have demonstrated an unprecedented ability to synthesize high-fidelity visual sequences. However, a fundamental gap persists between visually plausible video generation and the functional requirements of a world model, particularly in maintaining a stable and reasonable internal state over extended temporal horizons. While existing benchmarks primarily emphasize visual quality, motion coherence, and text-video alignment, they largely overlook memory, the core capability of a world model to preserve consistency across long-term horizons and complex interactions. To address this gap, we present \textbf{MBench}, a comprehensive benchmark dedicated to quantifying and evaluating the memory capability of video world models. We systematically decompose the memory capability of video world models into three hierarchical and complementary core dimensions: entity consistency, environment consistency, and causal consistency, which are further refined into 12 quantifiable sub-dimensions for comprehensive characterization of long-term memory. Our benchmark is built upon rigorously curated real-captured long videos, and evaluated by rule-based quantitative matrices and VLM to enable objective and comprehensive consistency assessment. Extensive evaluations of mainstream state-of-the-art video world models reveal critical systemic limitations of existing methods in long-term state retention, providing a standardized benchmark and clear research direction to advance the field.

World models, internal simulators that learn the structure and dynamics of an environment, have emerged as a central paradigm in the pursuit of artificial general intelligence, enabling agents to predict, plan, and reason within learned representations. Despite rapid progress across reinforcement learning, robotics, autonomous driving, and video generation, the field lacks a unified framework integrating its diverse architectural choices, training methods, reasoning mechanisms, and application settings. This survey addresses that gap with a multi-axis taxonomy organized along four dimensions: (i) architecture, encompassing representation format, dynamics formulation, input modality, learning paradigm, and downstream application; (ii) methodological family, including state-space and recurrent approaches, transformer-based models, diffusion-based generators, physics-informed networks, and language-augmented multimodal systems; (iii) reasoning strategy, covering imagination-based planning, latent policy learning, counterfactual reasoning, and planning under uncertainty; and (iv) application domain, spanning robotics, autonomous driving, video prediction, multimodal agents, reinforcement learning, scientific modeling, medical imaging, educational measurement, and business and finance. Tracing the field from early cognitive-science foundations to milestone systems such as PlaNet, the Dreamer family, MuZero, Sora, Cosmos, and Genie, we examine how these dimensions interact and highlight the recent convergence of chain-of-thought reasoning with world-model imagination. We review evaluation protocols and benchmarks, identify persistent challenges such as compounding prediction errors, sim-to-real transfer, and fragmented evaluation, and outline future directions toward unified multimodal world models, foundation-scale interactive simulators, and safe deployment in safety-critical domains.

World models for embodied AI must be physically viable: constructed to answer intervention queries by representing the physical structure governing action outcomes, rather than merely predicting future observations. Existing observation-predictive world models can produce visually plausible but physically wrong rollouts. This failure is structural; distinct physical systems can look identical yet diverge under intervention. We expose this problem with controlled benchmarks that fix the visible scene while varying latent physics. We show that such models may recommend infeasible actions, mispredict interaction outcomes, or certify unsafe behavior. We argue that embodied AI requires world models that identify the simplest physical abstraction sufficient to answer an intervention query. Such a model comprises modular components, including environment representation, latent state and parameter estimation, action specification, interventional dynamics, and query-level response. An autonomous orchestrator should identify the relevant abstraction and compose compatible learned and structured components per query. When closed-form physics is unavailable, uncertain, or costly, the transition model may be analytic, simulated, learned, or hybrid, but it must preserve the structure that determines interventional outcomes. This decomposition makes the model interpretable, its components verifiable, and its outputs auditable against the query. It also provides a design principle for new world models and a feasibility test for existing ones: the right abstraction is not the most detailed model of the world, but the simplest model that preserves the distinctions relevant to the query. We demonstrate this approach on queries that existing systems fail to answer correctly, and outline how an orchestrator can dynamically assemble and adapt physically viable models for planning, control, and verification.

World models require state tracking, which is the ability to maintain a correct latent state across action sequences. Existing benchmarks are often synthetic or language-based, limiting their value as tests of structured state updates in realistic domains. We introduce Chess-World-Model, a large-scale state-tracking benchmark built from 10 million real chess games, where models predict the exact board state reached after a sequence of legal moves. Alongside a held-out real-game split, we include an out-of-distribution split from uniformly random legal play, which tests whether models learn the transition rules rather than shortcuts from common human positions. Prior theoretical and empirical work has shown that Transformers struggle to state-track, while input-dependent linear RNNs require expressive state-transition matrices to do so. We therefore benchmark a causal Transformer, block-diagonal SLiCE, Mamba-3, and Gated DeltaNet with negative eigenvalues under a matched interface and training protocol. The recurrent models strongly outperform the Transformer at 3 and 8 million parameters. Real-game performance saturates above 18 million parameters, but the random-uniform split remains discriminative up to 40 million, exposing failures otherwise hidden by scale. Additionally, ablations show that less expressive state-transition mechanisms reduce performance on the out-of-distribution split for all three recurrent models. Together, these results establish Chess-World-Model as a practical large-scale benchmark for state tracking that exposes failures model scale would otherwise conceal.

Action-conditioned world models are increasingly used as scalable simulators for robot learning, yet current evaluations provide limited evidence that their predictions are reliable under the actions they condition on. Existing benchmarks largely emphasize visual fidelity, leaving unclear whether predicted futures are physically plausible, faithful to commanded actions, and calibrated to failure when actions should not succeed. We introduce \textsc{MiraBench}, a hierarchical benchmark that defines \emph{action-conditioned reliability} as a core evaluation target for robotic world models. MiraBench decomposes this target into three progressively demanding levels: \emph{Physics Adherence}, which evaluates reference-free physical consistency; \emph{Action-Following Fidelity}, which measures whether predictions respect task-relevant action inputs; and \emph{Optimism Bias Detection}, which probes the tendency to predict successful outcomes under failure-inducing actions. To support this evaluation, we curate a human-annotated corpus with over 16,000 judgments across tasks, failure categories, and leading world models. We evaluate 12 representative model configurations spanning vector-conditioned robotic world models, text-conditioned generative world models, open-weight systems, closed-source systems, and multiple model scales. Across this broad model landscape, MiraBench reveals three central findings: visual fidelity is a poor proxy for action fidelity; increasing model scale does not reliably improve action following; and optimism bias is pervasive across current systems. By shifting evaluation from appearance to action-conditioned reliability, MiraBench provides a diagnostic foundation for assessing and improving robotic world models as faithful simulators.

Robotic manipulation depends on the ability to anticipate how actions reshape objects, contacts, and scene geometry before execution. Learned world models provide this capability by predicting task-relevant future evolution under robot intervention, yet the term now spans latent dynamics models, action-conditioned video generators, three- and four-dimensional scene predictors, physics-informed simulators, and predictive modules inside vision-language-action systems. This breadth has fragmented the literature and obscured the design choices that matter for manipulation. We survey world models for robotic manipulation through three questions: what future representation is predicted, how prediction is connected to action, and when prediction is used in the robot-learning pipeline. We operationally define a world model as an action-conditioned predictive system and distinguish it from perception modules, inverse models, policies, rewards, and value functions. We then organize existing work into five representation families, develop a functional taxonomy that separates integrated prediction-action models from explicit predictive planners, and characterize infrastructure roles including synthetic experience generation, candidate filtering, search-based evaluation, learned environments, and outcome verification. We further map these roles across pretraining, post-training, and inference adaptation, review 34 manipulation datasets, and synthesize evaluation protocols for predictive fidelity, task performance, and simulator reliability. This survey shows that world models are evolving from task-specific dynamics predictors into predictive infrastructure for robot learning, while exposing open challenges in contact modeling, hallucination control, action alignment, and benchmarking under closed-loop use.

Video generation models are increasingly used as world simulators for tasks like driving and robotic manipulation. What matters in these settings is not whether a single video looks right, but whether the model's output changes when its input changes. We test this by giving a model two prompts describing the same scene with one physical detail varied, and checking whether the two videos diverge the way physics predicts. The wording difference between the prompts is small by design, since only one variable is changed, but the correct physical difference is not. A model that misses this can still produce two videos that each look plausible individually, and existing benchmarks score videos one at a time and cannot detect this failure. We introduce What-If World, 319 such prompt pairs built on real frames from nuScenes and DROID, organized by a taxonomy of six physical variables shared across driving and manipulation. Each pair is scored with APEO, a four-part rubric checking whether each video follows its prompt (Adherence), is physically consistent (Physics), preserves the shared scene (Environment), and ends in the correct difference (Outcome). Across nine state-of-the-art models, no system exceeds 52% on the paired score, and open-source models cluster near 28%. Every model tested fails on a large fraction of causal interventions, indicating substantial room before these models can reliably support action-conditioned simulation or model-based planning. Where models do score well, performance appears to track the visual prominence of the intervention rather than the tractability of its underlying physics. Some visually subtle interventions score as low as 14.2%, while visually pronounced ones reach 40.4%.

Interactive world models are advancing rapidly, yet existing benchmarks cover only part of the required competencies, leaving no unified standard for systematic evaluation. To fill this gap, we introduce WBench, a comprehensive multi-turn benchmark for interactive world model evaluation along five dimensions, namely video quality, setting adherence, interaction adherence, consistency, and physics compliance. WBench contains 289 test cases and 1,058 interaction turns, where each case specifies a world setting and a multi-turn interaction sequence, covering diverse scenes, styles, subjects, and both first- and third-person perspectives, together with four interaction types, including navigation, subject action, event editing, and perspective switching. For navigation, WBench unifies text, 6-DoF pose, and discrete-action control, enabling evaluation of models with different native input interfaces. Evaluation uses 22 automatic sub-metrics that combine specialist vision models with large multimodal models, and all metrics are validated against human judgments. Across 20 state-of-the-art models, we find that no single model performs strongly across all dimensions. We provide detailed diagnostic insights into the characteristic strengths, weaknesses, and open challenges of each model. Code and data are available at https://github.com/meituan-longcat/WBench.

Video world models have achieved strong visual realism, but this does not ensure that their dynamics are truly governed by actions. In this work, we argue that action faithfulness should be understood through the compositional structure of actions, which in many embodied settings follows a group structure (e.g., SE(2) for navigation). Based on this insight, we formalize action-conditioned world modeling as realizing a group action on the state space, providing a principled criterion for evaluating dynamics beyond visual quality. To operationalize this framework, we propose a unified approach that enforces identity, inverse, and composition consistency via latent-space regularization with synthesized supervision, avoiding additional data collection. We further introduce two metrics: Group-Action Consistency (GAC) and Group-Action Robustness (GAR), to evaluate structural correctness and rollout stability. Extensive experimental results show that our method consistently improves both GAC and GAR in state-of-the-art video world models without degrading perceptual quality.

Despite the growing use of world models as decision-making agents, their adversarial robustness remains underexplored due to the lack of dedicated automated evaluation methods. A key obstacle is that attack evaluation must be both accurate and efficient: weak manually tuned attacks can overestimate robustness, while exhaustive hyperparameter search is prohibitively expensive because each candidate requires closed-loop rollouts through learned latent dynamics. We introduce WMAttack, an automated attack-search framework for adversarial evaluation of world-model agents. WMAttack formulates robustness evaluation as a finite-budget search over attack configurations, including attack families, perturbation budgets, optimization steps, restarts, and allocation rules. To improve search accuracy, Self-Correcting Attack Search (SCAS) refines the attack proposal distribution using feedback from reward degradation, action instability, runtime cost, and rollout variability. To improve search efficiency, Representation-Guided Attack Retrieval (RGAR) retrieves effective historical configurations from representation-similar tasks, providing a warm start for unseen environments. We provide a theoretical explanation showing that proposal refinement improves finite-budget search when it shifts probability mass toward high-utility attacks. Across Atari and DeepMind Control tasks, WMAttack consistently discovers stronger attacks than the evaluated baselines, improving normalized reward drop from 0.497 to 1.034 on DreamerV3 Atari and from 0.319 to 0.682 on DMC. Ablations further show that RGAR improves initial candidate quality and SCAS improves final attack utility under fixed evaluation budgets.